Work
Sahil Kadadekar
Founding ML engineer building production agentic and inference systems for clinical AI, cybersecurity, and model deployment.
Architected Attunica's AWS clinical platform, live in 2 pilots including a 120-therapist clinic; at GhostEye (YC S25), shipped Beacon to 5 enterprise pilots, cutting conversation LLM costs 30–80%. Built Chimera, a six-subsystem constitutional AI platform backed by 55 reports / 1.46M+ measurements, 2 workshop-accepted papers (the ICML 2026 Workshop on Hypothesis Testing and the NeurIPS 2026 Workshop on Foundation and Large Model Security), 1 more under double-blind review, 2 public arXiv preprints, six merged upstream fixes, 23 Hugging Face models, and two PyPI tools with 51K+ downloads.
PICT + Cross-Institutional Research Collaborations Research Engineer — BCI, EEG & Medical Imaging Nov 2020 – Sep 2023
Stealth Startup in Medical AI Co-Founder Oct 2023 – Aug 2025
New York University Graduate Project — Visual Anomaly Detection Sep 2024 – Dec 2024
Chimera Founder & Lead ML Architect Sep 2025 – Present
Attunica, LLC Co-Founder & Head of Engineering Oct 2025 – Present
GhostEye Inc. (YC S25) Founding Machine Learning Engineer Dec 2025 – Mar 2026
Research & Open Source
ICML 2026 Agent Reproducibility Challenge
48 papers · 118 claims · 297 pts, #26 of 1,221 (top 2.1%)
- Built an autonomous, provider-neutral producer/reviewer/root-coordinator pipeline under per-paper evidence contracts fixed before execution; no human-intervention loops, no paper code reused, with every command, exit code, and output hash logged. A 4-paper concurrent calibration wave reached independent review in 27m42s; challenge referees scored each runnable Hugging Face Space.
Show 1 moreShow fewer about ICML 2026 Agent Reproducibility Challenge
- Two of the 3 official falsifications were Hierarchical Successor Representation claims, the third a nonparametric-regression re-calibration; my scaled DropoutTS rerun also failed its +46% robustness headline (independently corroborated; referee-scored toy), while spectral-bound theory reproduced to machine precision (1e-16 residuals).
LLM Safety Research & Evaluation Infrastructure
55 reports / 1.46M+ measurements · 2 accepted + 1 under review + 2 preprints
- Built the Banterhearts execution substrate: shared multi-backend evaluation and serving harnesses (Transformers, Ollama, ONNX, vLLM, SGLang, TGI), per-sample JSONL provenance, seed/config/git manifests, checkpointed OpenAI/Anthropic batch judges, disagreement-aware triangulation, fail-closed analyzers, and frozen-byte paper packages under dependency-locked CI.
Show 7 moreShow fewer about LLM Safety Research & Evaluation Infrastructure
- Led the independent program across training, deployment, and inference: consumer-GPU discovery with bounded A100 confirmation; pre-registered paired designs, bootstrap CIs, TOST, and Holm-Bonferroni. Presented at the ICML 2026 Workshop on Hypothesis Testing; accepted at the NeurIPS 2026 Workshop on Foundation and Large Model Security (safety-embedding directions, the first paper on Chimera’s own thesis); 1 more paper under double-blind review; two studies public on arXiv, on quantization safety and temperature-zero speculative decoding. Every paper is first-author.
- Tested three serving-performance assumptions: an M/D/1 model with linear parallel-service scaling underestimated queue wait by up to 20.4×; Ollama NUM_PARALLEL changes showed no significant effect in 30 contrasts; direct PyTorch lost more per-agent throughput than Ollama as concurrency rose from one to eight (86.4% versus 82.1%). A separate N=8 comparison found up to 2.25× vLLM throughput over Ollama.
- Isolated deployment effects with a 78,183-row, four-stack ablation; separate safety experiments found chat-template divergence could exceed numerical-precision effects. Designed controls for backend, template, and concurrency identity so changes in the serving environment were not misattributed to model weights; deployment rules include Q4_K_M and compile-prefill-only on Linux.
- The TAIS preprint examines output differences and refusal behavior under temperature-zero speculative decoding.
- Across two shared anchor models, normalized safety-score changes split 57% quantization, 41% backend, and 2% concurrency: descriptive shares, not a causal decomposition. Across 18 models / 10+ families (15 scored), alignment type showed no statistically detectable association with fragility (p=0.942) and 4 mechanistic probes failed to predict it, while output instability was strongest (r=0.909).
- Showed safety can degrade 13.9× faster than quality under quantization and found hidden refusal degradation in 7 of 11 AWQ/GPTQ conditions; the Refusal Template Stability Index (RTSI) triages them for direct testing. Shipped RTSI + JTP and RTSI-gated routing, recovering 76% of the refusal gap by routing the riskiest 20% of configurations (LOOCV AUC 0.84).
- Isolated FP8 KV-cache precision in 24,054 paired records, then replicated the null on 7,578 records / 12 of 12 TOST-equivalent cells.
Chimeraforge — LLM deployment planner
v0.34.0 · 34,000+ downloads · 2,066 tests
- A 13-command CLI/Python/MCP planner (five MCP tools) that ingests JSONL or live vLLM/SGLang telemetry; searches model × quantization × backend × GPU/TP/PP plans across heterogeneous fleets; propagates weakest-source evidence; predicts VRAM, TTFT/TPOT, throughput, KV-cache/offload, prefix caching, multi-LoRA, cost, and energy; and emits vLLM/TGI/SGLang/Ollama launch commands, API break-even, and provenance briefs.
Show 1 moreShow fewer about Chimeraforge — LLM deployment planner
- Made its evidence hierarchy executable: every estimate is labeled measured, extrapolated, derived, estimated, or unknown; validation separates in-corpus lookup from out-of-sample estimates, and --expect-fingerprint aborts if the pre-registered matrix changes. Validated on within-registry record-level holdouts: VRAM R²=.968, throughput R²=.859, composite quality RMSE=.062 (0–1 scale), latency MAPE=1.05%; stale prices and unsupported estimates fail closed.
EvidenceMCP Registryquantfit — quantization safety measurement CLI
v0.12.16 · 17,000+ downloads · 1,369 tests
- Implements QSR spec v0, a versioned quantization-safety measurement protocol, across AWQ/GPTQ/SmoothQuant/FP8/RTN/GGUF with capacity preflight; its two-axis release gate (refusal robustness + over-refusal) uses at-risk denominators, Wilson CIs/power and minimum-detectable-effect checks, revision-pinned artifacts, exact-engine provenance, stable JSON/exit codes, and JUnit where unmeasured axes skip rather than pass.
Show 1 moreShow fewer about quantfit — quantization safety measurement CLI
- Completed a 15-target screen (14 measured; 0 dangerous-axis regressions across 12 GGUF + 2 compressed-tensor targets) with a passing sensitivity control; calibrated and replaced a judge with 56.2% false positives, then human-adjudicated 11 flags (6 real, 5 judge errors), preventing an approximately 2× overclaim. Cross-hardware T0 replications invalidated an apparent safety breach, so I voided the result rather than publish it.
Hugging Face — 23 model releases
- Published 23 Hugging Face models: 11 AWQ/GPTQ 4-bit + 6 FP8-Dynamic releases across Llama 3.2, Qwen 2.5, Mistral, Gemma 2, and Phi-2; 4 GPT-2 scaling variants; the MedMCQA Dr.GRPO LoRA adapter; and a ModernBERT refusal classifier (0.9773 macro F1 on 441 unambiguous XSTest GPT-4 responses, trained on WildGuardMix).
QuantSafe Certifier
- Shipped the QuantSafe Certifier (≤32B): 4-delta refusal screen, semantic cross-check, multi-judge stack, constitutional debate, and Ed25519-signed certificates.
vLLM PR #45207 — merged
- Merged vLLM PR #45207 (benchislett-approved, merge 55da232): fixed a KV-cache page-size unification crash for hybrid Mamba/attention models by preserving Mamba block granularity while padding pages through page_size_padded; regression test added (fixes #43626), then independently reproduced on NVFP4 + DFlash speculative serving.
EvidenceFixes vLLM #43626PyTorch PR #175562 — merged
- Landed upstream via PyTorch PR #175562 (jansel-approved, be90a14, shipped in 2.13 stable): torch.compile / CUDAGraph trees dealloc hardening against diagnostic-metadata divergence + CUDA regression test; isolated compiled-decode failure #175557 and validated maintainer PR #184102.
PyTorch PR #190555 — merged
- Merged PyTorch PR #190555 (jansel-approved, 0b96f88; Inductor): split cross-device extern kernels out of CUDA-graph partitions, which were capturing CPU storage and failing memory-pool checks; same-device kernels stay graph-eligible, with regressions for custom ops, multi-output ops, index_put, and SDPA dropout.
PyTorch PR #199075 — merged
- Merged PyTorch PR #199075 (guilhermeleobas-approved, 0055968; Dynamo): Python random float draws were traced as float32 while Inductor's CPU kernels received float64 buffers, so x * rng.random() returned values around -9e34; the graph input now uses the same full-precision conversion as runtime, with an Inductor regression test (fixes #198187).
EvidenceFixes PyTorch #198187Ollama PR #16669 — merged
- Merged Ollama PR #16669 (dhiltgen-approved, commit fc58544): root-caused two enumeration bugs in Ollama's Go discovery path causing inverted iGPU/dGPU Vulkan classification on Windows hybrid graphics; about 9× faster inference on the affected system, with regression tests for both failure modes.
Triton PR #10819 — merged
- Merged Triton PR #10819 (peterbell10-approved, b92dc43): fixed a tl.flip compile-time crash on the documented default dim=None — resolve the dim before the bounds assert + add a dim=None test (fixes #10790).
Earlier publications
IJRASET 2023 · JETIR 2022
- Digital Currency Price Prediction using Machine Learning. Sahil Kadadekar, Sumaiya Shaikh, Isheeta Shahir, Hemantkumar Mali, and Rupesh Jaiswal. IJRASET 11(9):338–355, 2023.
Show 1 moreShow fewer about Earlier publications
- Machine Learning Based Car Damage Identification. Mansi Satpute, Sahil Kadadekar, and Rupesh C. Jaiswal. JETIR 9(10):b684–b690, 2022.
EvidenceJETIR 2022 (PDF)
Projects
Interactive builds, each running live on its own page.
Flight Routing Lab
With a tight deadline, planning ahead gets 40 of 64 simulated trips to JFK on time where booking the earliest nonstop gets 0, but 8 fewer arrive at all.
Offline Policy Evaluation Lab
A new decision policy looks 0.27 better on records of past decisions, but a control that ignores the situation gets 65% of that, and a changed reward erases it.
Customer Service Environment
Two scripted runs refund the same $48.00 on a double-charged order: refunding the duplicate charge scores 1.00, refunding the real one scores −0.40.
Code Verification Lab
Four candidate bug fixes: a two-test smoke suite passes three, the full six-test suite passes one. Passing the bug’s own example is not a fix.
Spreadsheet Reasoning Inspector
Rules vote on which spreadsheet values are final: balanced voting marks a scratch cell final, and a stricter gate finds neither real result.
Browser Agent Completion Gate
Six attempts at one booking: the success notice claims three, my earlier agent’s check accepts five, and only one saved the booking asked for.
Intake Triage Scorer
All 12,672 signal combinations scored: shouting “URGENT” never moves a message’s priority, and once a safety gate fires, nothing else does.
Staged Document Search
When a filtered search finds too few notes it drops filters: at its default it drops all three, says “ready”, and 2 of its 4 notes break the request.
Send Pacing Scheduler
In all 1,000 runs of its own example, the scheduler sends its last two messages before a person could have typed them.
Drone Mission Governance
With no telemetry at all the drone flies all 6 waypoints, and the mission’s lost-link instruction is never read; stale data mid-flight is only logged.
Collaborative Whiteboard
Two people recolour one note at once: in 6 of the 8 possible orderings a screen shows a colour the database lacks; numbered order fixes all 8.
Experience
Dec 2025 – Mar 2026
Founding Machine Learning Engineer
GhostEye Inc. (YC S25) · New York, USA
- Built Beacon, GhostEye's multi-agent JIT security training, in 90 days across web, Slack/Teams, SMS/RCS, WhatsApp, Telegram, voice, and email; 5 enterprise pilots (a top-10 global asset manager, a Fortune-100 cloud platform, Eight Sleep, Fella Health, ZeroPath). The asset manager pilot cut phishing click rate 58% and tripled reporting within one quarter; a shared JIT agent personalized remediation from vectorized phishing/vishing/smishing/deepfake failure histories.
- Fine-tuned self-hosted Llama-3-70B for phishing simulation (domain adaptation) via LoRA/QLoRA + DeepSpeed on 1M+ NIST-grounded emails (vendor impersonation, typosquat logins). Built Go orchestration for credential-harvest pages/session-cookie capture, traced scoring/adversarial logs, Azure tenant auth, country-code routing, STT/TTS fallback, and SCORM/Vanta reporting aligned to SOC 2/NIST/ISO 27001/GDPR.
Show 3 moreShow fewer about Founding Machine Learning Engineer, GhostEye Inc. (YC S25)
- Cut deepfake phishing simulation streaming response latency from about 40s in early benchmarks to 100–450 ms by rebuilding the pipeline as a multi-agent WebRTC system spanning synchronized video rendering, voice generation, human-like scheduling, and retry-aware delivery.
- Built analytics/evals for Vapi voice, self-hosted SMS simulation, and the training pipeline (TTFT, tokens/query, tokens/sec; in-house LangGraph evals, Kafka/ClickHouse scheduling and logging); cut per-turn latency from 5–7s to 0.5–1.5s with barge-in, streaming, caching, and summarization; caching and summarization cut conversation LLM costs 30–80%, with larger savings on longer conversations.
- Built C-suite OSINT knowledge graphs for enterprise clients incl. Fortune-100 (Firecrawl into Amazon Neptune, TinkerPop/Gremlin; 5K+ nodes per graph); fixed Neptune overload by restructuring around cross-executive overlap and per-executive connectivity; cut amortized render from ~360 to ~60 ms per node (83% lower) via caching and engagement-ranked level-of-detail loading.
Oct 2025 – Present
Co-Founder & Head of Engineering
Attunica, LLC · New York, USA
- Architected and solo-built Attunica's multimodal psychotherapy-training and evaluation platform: real-time LiveKit + Gemini + Anam sessions, durable transcripts/debriefs, and academic workflows (instructor assessments, verified recording playback); now lead a PM + 2 engineers; shipped to 2 pilots (NYU Silver MSW program; a 120-therapist clinic); HIPAA BAAs with Anthropic + AWS.
- Built a consent-gated, clinician-graded LLM-persona environment for measuring simulated humanity and attachment without patient data: persona text is character data, never instructions; tenant RBAC and signed BFF capabilities fence roles; immutable eval IDs, revocable consent, private versioned media, and two-channel Deepgram bind evidence to provider identity.
Show 3 moreShow fewer about Co-Founder & Head of Engineering, Attunica, LLC
- Rebuilt session scoring as three-sample median judging over alias-verified therapist evidence (wrong-speaker citations fail closed; input 32k → 15–18k tokens); a 40-fixture judge-validity benchmark with held-out scenarios and bias probes showed debiasing moved leniency from +0.81 to within 0.13 anchors of zero (held-out weighted kappa 0.66 vs 0.05 null; model-written key) and that sample disagreement alone misses large errors; eval records distinguish no evidence from scored zero; persisted provider-attempt state separates pre-call failures from ambiguous outcomes, preventing replay of billable side effects.
- Led the AWS-funded production cutover with Avahi to ECS, Aurora PostgreSQL 18, and Bedrock (retired Fly/Neon/Vercel) through a 141-row production-readiness matrix in which a PR cannot weaken the checks that approve it; qualified the release candidate with PostgreSQL 18 migration rehearsals and commit-bound artifacts under IAM-scoped credentials, zero SDK retries, bounded timeouts, and no cross-provider fallback or replay.
- Lead the Article 31 documentation product on AWS ECS (clinician-reviewed Claude notes/treatment plans): release-only deploys with drained, snapshotted in-VPC migrations; PHI-blind admins and psychotherapy-note authorship enforced end to end; a fail-safe restricted-role FORCE-RLS rollout; versioned clinical writes with append-only audit; on-device Whisper.
Sep 2025 – Present
Founder & Lead ML Architect
Chimera · New York, USA
- Architected a six-subsystem constitutional AI platform (2.3K+ core tests) linking a multi-provider voice/tool JARVIS gateway, calibrated router, multi-model debate, RLAIF, Rust provenance runtime, and Muse over pinned HTTP/JSON contracts. Requests traverse verifier → judge → debate → enforcement/canary/ZK; debate outputs feed retraining under drift and rollback gates.
- Falsified one-class safe-centroid routing across 3 corpora / 4 encoders (AUC 0.358–0.545), traced the failure to topic confounding, and replaced it with a supervised safe-minus-unsafe direction learned from labeled and debate pairs behind calibrated fast-path, debate fallback, canaries, and rollback.
Show 4 moreShow fewer about Founder & Lead ML Architect, Chimera
- Selected objectives by evidence shape: generated paired debate preferences for a self-hosted Llama-3-70B student and trained/released DPO-aligned models; implemented and one-step validated ORPO, Dr.GRPO, RLOO, and REINFORCE++ trainer paths; designed KTO for unpaired constitutional verdicts under shared evaluation and promotion gates.
- Diagnosed zero reward variance in 76.3% of rollout groups in a preregistered Dr.GRPO + LoRA RLVR study on MedMCQA. After revising the environment prompt and remeasuring the pre-RL baseline, a second run improved held-out pass@1 from 40.6% to 49.4% (+8.8 points, n=500; McNemar p=0.00034). Released both experiments and the adapter.
- Built a Rust alignment runtime across 7 crates: BFT consensus, Ed25519 provenance, Merkle verification, Ristretto255 Pedersen/Schnorr ZK proofs, Arrow IPC, sandboxed tools, and CRDT state sync; every governed decision emits a signed, replayable trace.
- Built JARVIS's governed temporal memory with a typed replay op-log, live write/recall, encryption, DSR/erasure fences, approval-gated destructive actions, and 5 channel adapters; external LongMemEval/LoCoMo gates shipped chunked retrieval (+7.31pp R@1) and timestamp evidence (+30.67pp LoCoMo-WHEN) while rejecting regressions. Muse converts OTel/ClickHouse traces into signed watcher → triager → fixer remediation records.
Oct 2023 – Aug 2025
Co-Founder
Stealth Startup in Medical AI · New York, USA
- Led 3 engineers + 1 clinician across 5 institutions building an ML-guided diagnostic platform (3 clinical domains, 1K+ cases) on AWS (Lambda, SageMaker, Bedrock, SQS/DLQs) + Qdrant RAG under HIPAA; 75% lower infrastructure cost.
Sep 2024 – Dec 2024
Graduate Project — Visual Anomaly Detection
New York University · New York, USA
- Implemented and benchmarked PatchCore and FastFlow on MVTec-AD using per-category AUROC; built a FastAPI service with Qdrant retrieval to return the five most visually similar anomalies for inspection.
Nov 2020 – Sep 2023
Research Engineer — BCI, EEG & Medical Imaging
PICT + Cross-Institutional Research Collaborations · Pune, India
- Adapted a channel-fused Dense CNN for BCI Competition IV-2a four-class motor imagery from 22-channel, 250-Hz EEG; deployed real-time inference in a physical NVIDIA Jetson prototype, improved preprocessing throughput 35% via pinned memory and asynchronous CPU–GPU transfers, and earned PICT Honors in AI & ML through the work.
- Studied five-class diabetic-retinopathy grading over 92K+ fundus scans (1.2 TB), comparing about ten attention mechanisms with clinician-facing SHAP analyses; sharded loading and multi-GPU training improved training throughput 4×.
Show 1 moreShow fewer about Research Engineer — BCI, EEG & Medical Imaging, PICT + Cross-Institutional Research Collaborations
- Built DenseNet201 dental classification over 1K+ expert-annotated cases and 40K+ images; dataset and annotation protocols were adopted by 5+ research teams and received institutional copyright L-122721/2023.
Service
2026
NeurIPS 2026 reviewer
Workshops: FLMSec (2), JUDGe (3), RTCA (5). Ethics: main conference (1), Evaluations & Datasets track (1). All 12 reviews completed.
2026
Journal reviewer
Advances in Artificial Intelligence and Machine Learning (AAIML; ISSN 2582-9793), a Scopus- and Web of Science-indexed journal. One review completed.
2026
Hackathon judge
Build for the Border (May 2026) and AI Healthcare Hack NYC (Jul 2026).
Education
New York University — Tandon School of Engineering
New York, USA
M.S. in Computer Science
GPA: 3.5/4.0May 2025
Pune Institute of Computer Technology
Pune, India
B.E. in Electronics & Telecommunications, Honors in AI & ML
CGPA: 9.1/10.0Aug 2022
Technical Skills
- LLM systems
- PyTorch, Transformers, DeepSpeed, Accelerate, Ray, vLLM, SGLang, TGI, TensorRT-LLM, llama.cpp/GGUF, continuous batching, KV-cache optimization, speculative decoding
- Inference optimization
- CUDA, Vulkan, Triton, TensorRT, FlashAttention, ONNX Runtime, torch.compile, Nsight Systems/Compute, NVIDIA DALI, GPTQ, AWQ, FP8, INT4/INT8
- Post-training & evals
- LoRA/QLoRA, DPO, ORPO, KTO, GRPO, Dr.GRPO, DAPO, RLOO, REINFORCE++, RLAIF, PRM/ORM routing, WARM judges, TOST equivalence, Holm-Bonferroni, Cohen's d, LOOCV, SHAP, SciPy
- Product stack
- Python, Go, Rust, TypeScript, SQL, C++, C#, FastAPI, Next.js, React, PostgreSQL/pgvector, Redis, ClickHouse, Qdrant, DynamoDB, Docker, Kubernetes, AWS (ECS, Bedrock, Aurora), Azure
- Voice & agents
- LangGraph, LangSmith, MCP, LiveKit, Gemini Realtime, Whisper STT, TTS, WebRTC streaming, multi-provider gateways