I build reliable software for teams that need to control AI costs, automate data workflows, and make high-stakes decisions auditable.
- AI cost and policy control: model routing, budget enforcement, fallback rules, guardrails, usage attribution, and audit evidence.
- High-integrity backend systems: deterministic decision services, replayable workflows, tamper-evident logs, concurrency-safe accounting, and low-latency Rust APIs.
- Data engineering and automation: scraping pipelines, CSV/Excel cleanup, validation contracts, scheduled jobs, API integrations, and client-ready reports.
- Decision-focused machine learning: calibrated risk scores, uplift modeling, explainability, counterfactuals, and ranked operational actions.
I am available for focused freelance and contract work involving:
| Engagement | Typical delivery |
|---|---|
| Rust backend and systems engineering | Performance-sensitive APIs, concurrency-safe services, deterministic kernels, verification tooling, and production hardening |
| AI infrastructure | Model gateways, RAG and retrieval services, provider routing, cost governance, evaluation, observability, and policy enforcement |
| Python data products | FastAPI services, validated ETL pipelines, web scraping, CSV/Excel transformation, scheduled automation, and reporting |
| Applied ML systems | Churn, lead scoring, pricing, uplift, forecasting, calibration, SHAP, and decision queues |
| Technical audit and rescue | Architecture review, correctness testing, performance diagnosis, CI repair, security boundaries, and maintainability upgrades |
Every engagement is scoped around a measurable outcome, explicit failure modes, tested deliverables, and a maintainable handoff.
Calybris Core is a deterministic, proof-carrying decision kernel for routing, guardrails, and budget-sensitive automation.
catalog + policy + request → decision + verifiable audit bundle
It is designed for systems where a team must be able to answer:
- Why was this model, provider, offer, or action selected?
- Which policy and budget state were used?
- Can the decision be replayed and independently verified?
- Can concurrent requests overspend or violate exposure limits?
Engineering evidence:
- Integer-only Rust hot path at approximately 115 ns per decision
- Byte-exact proof contract, golden vectors, replay verification, and
calybris-verifyauditor CLI - SHA-256 digests, hash-chained WAL with optional HMAC, and Ed25519 policy provenance
- CAS budget accounting with conservation invariants verified using Loom and Miri
- No hosted dependency and no
unsafein project code
cargo add calybris-core| Project | Business problem | What it demonstrates |
|---|---|---|
| ProofFrame | Data pipelines need fast validation and durable evidence of exactly what was checked | Arrow-native Rust/Python contracts, canonical fingerprints, keyed diffs, PII and leakage scans, signed proof receipts |
| Aegis | Churn scores are not useful unless teams know whom to contact and which action may help | Calibrated risk, uplift evidence, SHAP, counterfactuals, expected-value decisions, and an operations dashboard |
| Churn Prediction & Retention Report | Analysts need reproducible scoring and a stakeholder-ready action report | Validated ML pipeline, calibration, confidence intervals, model cards, ranked retention queue, and PDF delivery |
| CRM Lead List Cleaning API | Messy CRM imports create duplicates, invalid contacts, and unreliable automations | FastAPI cleanup service, email and phone normalization, deduplication, profiling, safe CSV export, Docker, and OpenAPI |
| Scrape Quality Pipeline | Scraped data must remain reliable when pages, selectors, or output schemas change | Async collection, polite rate limits, retries, typed records, Pandera validation, manifests, tests, and CSV/JSONL/Excel/Parquet export |
| Price Monitor Pipeline | Teams need repeatable public price checks instead of manual monitoring | Config-driven extraction, validated snapshots, threshold alerts, run manifests, and client-ready reports |
Clear scope → typed boundaries → tests and CI → observable execution
→ reproducible output → documentation and handoff
- Production-minded code with explicit assumptions and failure behavior
- Tests for correctness-critical paths and fixtures for external integrations
- CI, linting, dependency boundaries, and reproducible setup
- Honest evaluation: no inflated accuracy, ROI, performance, or AI claims
- Documentation aimed at the next engineer or operator, not only the original author
Languages: Rust, Python, TypeScript, SQL
Backend & data: Tokio, Axum, FastAPI, Litestar, Pydantic, Polars, Pandas, Arrow, DuckDB, PostgreSQL
AI & ML: model routing, RAG, XGBoost, LightGBM, CatBoost, SHAP, uplift modeling, contextual bandits
Delivery: Docker, Linux, GitHub Actions, OpenAPI, Typst, Next.js, React
Verification: property-based testing, Loom, Miri, replay tests, cryptographic digests, audit trails
If you are dealing with an unreliable data workflow, expensive AI pipeline, difficult Rust backend, or ML system that produces scores but not decisions, send me:
- the business problem,
- the current stack or data source,
- the expected deliverable,
- the target timeline.
I can help define a focused first milestone before expanding the scope.