Hossain Pazooki
The model interprets. A deterministic system executes.
I build data platforms, evaluation harnesses, and reliable agentic and LLM systems for regulated domains — point-in-time financial data, healthcare claims, cross-border compliance — where a wrong answer has a real cost and "it usually works" isn't good enough. open to software engineering roles in data platform, AI infrastructure, and platform reliability.
cross-model KV-cache transfer · agent evaluation under censored feedback
HIPAA-aligned healthcare RCM platform · team of 4
01Technical skills
grouped by where the work lands · core = daily, load-bearing · dashed = familiarSeven years writing, reviewing and debugging production code. Python and SQL are the daily instruments; the rest ship in the open-source work below.
Agents that propose; deterministic gates that dispose. Evaluation is the product, not the afterthought.
Instrumentation under the model, and the serving hooks that turn a fitted model into a production signal with drift you can see.
Event-driven services with durable state, idempotent consumers, and contracts generated from a single source of truth.
Point-in-time correctness and fail-closed data-quality gates: as-of day D returns only what was known by D.
The product surface, typed end to end from the API spec.
The platform, not the reviewer, carries the validation habits: required-check merge gates that have held a PR at BLOCKED under branch protection.
Least privilege, signed artifacts, and an audit record for every decision that can move value.
A check that passes either way proves nothing. Verdicts are written so a second reviewer can reproduce them.
Where the discipline started: the quantitative seat where a model that looks right and isn't moves capital.
02Experience
research · industry · founding · quant — most recent first- Cross-model KV-cache transfer. Model swaps (cascading, routing, mid-conversation handoff) discard the cache and repay full prefill. One model's cache is mapped into another's with a closed-form linear fit from a small calibration set — no gradients.
- Own cache and model instrumentation (PyTorch internals, HF attention hooks, GPU memory) and the mechanism research: experiment and intervention design, inference under small samples.
- Open question: at an agent handoff, can a short replay of the latest tool observation predict whether a translated cache changes the target's next action, without full prefill? Pair-level pre-fit screen plus a runtime commutation gate, fail-closed on disagreement.
- Companion track: evaluating LLM agents where feedback is selective or censored — planted ground truth, negative controls, fail-closed gates. Replication and harness in progress.
- Led a team of 4 building agentic compliance pipelines for healthcare revenue-cycle management: PydanticAI agents orchestrating claim review, denial prediction, and appeal generation as staged workers with human-in-the-loop checkpoints.
- Led incident response on a production calibration drift: root-caused systematic LLM over-confidence and rebuilt the decision layer as a 5-tier confidence-calibration and verification gate — every output mapped to ground truth, deterministic and LLM-as-judge evaluators, HITL escalation on uncertain calls, regression gates in CI. Contributed to a 34% reduction in claim denials.
- Built a hybrid retrieval pipeline (BM25 + vector search) over regulatory and payer policy documents, with a structured evaluation framework tracking precision, recall, and hallucination rates.
- Owned the platform a HIPAA-aligned org ran on: FastAPI microservices (50+ endpoints) on Kubernetes (EKS), PostgreSQL schema design and tuning, Redis, multi-source ETL of claims data, Terraform-managed AWS; CI/CD gated by 450+ tests with auto-rollback and strict typing. Set the team's code-review practice.
- Sat with billing specialists and compliance staff and turned their judgment into written evaluation guidelines and executable specs the team graded against.
- Sole engineer taking a full product 0→1 under ambiguity: architecture, CI/CD, and development practices from the first commit, owning every call from design through rollout.
- Designed the data model and core backend (Python, PostgreSQL) and shipped the React/TypeScript frontend; containerized AWS deployment, observability, releases end to end.
- Owned the ingestion and analytics pipelines feeding product and growth decisions.
- Built the ML platform for a metal additive-manufacturing R&D org, embedded with the scientists who used it daily: a unified data layer linking process parameters, characterization results, and production outcomes by part ID across sintering logs, mechanical-test data, and QA.
- Designed experiment tracking for design-of-experiments work, keeping iterations reproducible and comparable across weeks.
- Shipped model-serving hooks that scored new production runs automatically, plus monitoring that compared predictions against characterization ground truth to surface drift.
- Translated materials-team objectives — yield, faster material qualification — into tractable ML KPIs, co-designed with non-technical domain experts.
- Built marketing analytics models and reporting infrastructure for the revenue team, supporting data-driven ad sales and audience monetization across PMC's media portfolio.
- Led research on greenfield projects exploring new revenue streams in creative industries, reporting directly to the Director of Revenue.
- Econometrics research on sovereign wealth fund investment patterns, capital allocation models, and resource-based vs. pension fund return dynamics.
- Causal inference at scale: propensity/IPW, IV/2SLS, GMM, Heckman selection on cross-country panel data.
03Open-source projects
the same discipline, shipped in the open · statuses restated from each repoPlugin infrastructure other engineers install — 9 commands, 20 skills, 5 agents, 3 hooks, 12 check gates, marketplace install. A verification spine refutes each load-bearing claim (re-run the gate, dispatch a skeptic) before it is trusted; no claim credited until a negative control proves the check can fail. A git-guard hook refuses silent history rewrites, and a public ledger records its own misfires.
Measures model quality where the labels you need are the ones you can't observe: plant ground truth, hide it behind an approval policy, escalate severity until calibration correction fails. The deliverable is an operating frontier — 25 seeds, a distribution, not a point. A 13-claim gate recomputes every published number from hash-pinned CSVs and blocks publication when a pre-registered hypothesis fails — four did, and the claim was withdrawn. Written up in Towards AI.
The model never holds authority to act: an agent declares intent; a deterministic Go gate authorizes it against a human-signed, content-addressed policy (BLAKE3/ed25519, compiled in Rust) or refuses — unevaluable, duplicate, unsigned all deny. Every decision lands as one durable record carrying its trajectory hash, recomputable by a stdlib-only verifier that runs none of the gate's code. LangChain and MCP adapters ship with it; an NLI gate retracts unentailed rationale.
Security, observability and FinOps rules in OPA/Rego over Terraform and Kubernetes — 74 unit tests, 19 negative controls refused in CI, all 16 deny clauses proven load-bearing by mutation. Demonstrated blocking a merge: six required CI checks, no admin bypass — a planted violation turned exactly the job that reads it red (PR BLOCKED). Infracost cost-delta gating on pinned estimates, failing closed on a broken $0.00 parse.
As-of date D returns only what was filed and accepted by D — no lookahead across restatements. 181M bronze rows to 86.6M gold facts across 16,667 filers; publish gate proven able to fail on a planted twin (429,949 rows, exact), with an independent consumer gate red on that twin in CI.