I build, audit, and scale resilient cloud backends, high-throughput LLM proxy gateways, and custom Model Context Protocol (MCP) integrations.
Proven track record contributing high-reliability solutions to enterprise repositories with tens of thousands of stars.
πΌ Available for Contracts β’ π οΈ Case Studies β’ β‘ Tech Stack β’ π« Get in Touch
I believe code speaks louder than buzzwords. Here are real-world architectural and resilience contributions deployed to production open-source engines:
- Project:
BerriAI/litellm(25,000+ β β Enterprise LLM Proxy Gateway) - Pull Request:
#43888βfix(transport): dispose unclosed aiohttp ClientSession on transport __del__ - The Problem: High-concurrency client session eviction left underlying
aiohttp.ClientSessioninstances open without deterministic cleanup. In production clusters routing millions of tokens, unclosed sessions generated runtime resource leak warnings and risked socket exhaustion under heavy cache turnover. - The Solution: Engineered safe asynchronous event-loop inspection and deterministic session disposal on transport cleanup. Enforced strict lint budgets (
BLE001/S110compliance), eliminated non-deterministic sleeps with deadline polling, and structured full regression coverage in the mapped unit test suite. - Outcome: 40+ CI checks 100% green, verified zero connection leaks across continuous proxy load tests.
- Project:
BasedHardware/omi(20,000+ β β Open-Source AI Wearable & Ecosystem) - Pull Request:
#20016βfix(backend): validate inputs, clamp pagination boundaries, and guard keyset tombstones in mcp_conversation_pages - The Problem: In hosted Model Context Protocol (MCP) conversation card streams, encountering soft-deleted tombstones or corrupted documents lacking
created_attimestamps caused cursor position corruption (last_position = (None, doc.id)). The subsequent iteration threw an unhandledValueError, crashing caller endpoints with silent 500 errors instead of advancing pagination. Furthermore, negative limits/offsets bypassed boundaries straight into Firestore. - The Solution: Implemented non-destructive keyset cursor caching, boundary clamping on query limits/offsets, fast-return logic for inverted date ranges, path traversal sanitization, and 18 hermetic unit tests.
- Outcome: 18/18 hermetic tests passing, zero 500 crashes during MCP cursor pagination, official bounty proposal submitted.
- Project:
BasedHardware/omi(20,000+ β β Open-Source AI Wearable & Ecosystem) - Pull Request:
#20020βfix(backend): clamp non-negative costs, sanitize date paths, and harden llm_gateway_accounting - The Problem: In LLM gateway event accounting, negative cost values could decrement the organization's finops cost rollups via
firestore.Increment(-X), while float USD estimates were dropped to 0 due to rigidinttype checks. Furthermore, date strings containing slashes (/or\) caused accidental Firestore subcollection injection inllm_gateway_user_days, and unhandled plan resolution exceptions risked dropping valid billing attempt records. - The Solution: Engineered non-negative cost clamping, safe float/numeric string parsing with boolean rejection, path separator sanitization on date strings and attempt IDs, and fail-open exception boundaries on usage plan resolution.
- Outcome: 15/15 hermetic tests passing in 0.82s, 4/4 PR preflight checks passed, finops ledger integrity and single-collection index guarantees verified.
- Project:
BasedHardware/omi(20,000+ β β Open-Source AI Wearable & Ecosystem) - Pull Request:
#20023βfix(backend): validate identities, clamp index TTLs, and protect prefix keys in mcp_token_cache - The Problem: In hosted Model Context Protocol (MCP) OAuth token caching, missing fields in client identities caused unhandled
KeyErrorcrashes during valid session caching, while non-positive index TTLs caused Redis to immediately purge grant token indices. During grant revocation, empty/blank decoded token hashes produced the base prefix keymcp:oauth:at:, inadvertently wiping the base key namespace in Redis. - The Solution: Validated complete identity payload schemas, clamped grant index TTLs to positive values, guarded multi-token revocation against base key deletion, handled un-serializable payloads in HMAC integrity envelopes, and structured 41 hermetic unit tests.
- Outcome: 41/41 unit tests passing in 2.20s, zero unhandled
KeyErrorcrashes on MCP OAuth caching, Redis base prefix keys protected from accidental revocation wipes.
Available for freelance contracts, fractional engineering, and scoped sprints:
| Service | What I Deliver | Typical Engagement |
|---|---|---|
| LLM Gateway & Proxy Architecture | Custom routing, fallback cascades (Claude β Gemini β DeepSeek β Local Ollama), semantic caching, strict rate-limiting, and budget ratchets via LiteLLM / custom gateways. | 1β3 Week Sprint |
| Custom MCP Server Development | High-performance Model Context Protocol (MCP) tools, resources, and SSE streaming connectors for your proprietary APIs, databases, and internal workflows. | 3β10 Days |
| Backend Resilience & API Hardening | Concurrency contention mitigation, Firestore/PostgreSQL query optimization, Redis rate-limiting (Lua scripts), and eliminating unhandled 500 edge cases in Python/TypeScript. | Architecture Audit or Sprint |
| Zero-Downtime Migration & CI/CD | Automated pre-flight validation gates, Ruff/ESLint strict budgets, hermetic unit testing, and containerized Cloud Run / Docker pipelines. | Retainer or Project |
- Languages: Python (3.10β3.14), TypeScript, JavaScript, SQL, Bash.
- Frameworks & Gateways: FastAPI, LiteLLM, Express, Node.js, Next.js, Starlette, Pydantic (v2).
- Databases & Caching: Google Cloud Firestore, PostgreSQL, Redis (Pub/Sub & Lua scripting), Pinecone (Vector RAG).
- Architecture & Protocols: Model Context Protocol (MCP), SSE (Server-Sent Events), WebSockets, REST, gRPC, OAuth2 / OIDC.
- Testing & Quality Assurance: Pytest (Hermetic fakes & transactions), Vitest, Black, Ruff, ESLint, Playwright.
- DevOps & Cloud: Docker, Kubernetes, Google Cloud Run, GitHub Actions CI/CD workflows.
βββββββββββββββββββββββββββββββββββββββββββββββββ
β ENGAGEMENT MODELS β
βββββββββββββββββββββββββββββββββββββββββββββββββ€
β β‘ Milestone Sprint: $1,500 β $3,500 / task β
β π‘οΈ Fractional Monthly: $2,500 β $6,000 / mo β
β π Architecture Audit: $80 β $120 / hr β
βββββββββββββββββββββββββββββββββββββββββββββββββ
Looking to eliminate connection leaks, build custom Model Context Protocol (MCP) tools, or harden your LLM gateway before the next scaling milestone?
- Step 1: 15-Minute Architecture Teardown (Free)
- We audit your current LLM routing, latency bottlenecks, or MCP data requirements.
- π Schedule a 15-Minute Teardown on Cal.com
- Step 2: Scoped 5-Day Milestone Sprint ($1,500 β $3,500)
- Transparent, flat-rate pricing. Zero open-ended hourly billing.
- 50% upfront deposit, 50% upon clean pull request merge and test verification.
- Step 3: Verification & Zero-Downtime Handoff
- 100% hermetic unit test suites, zero CI/CD lint violations, and complete architectural documentation.
- Booking: cal.com/bangersoul
- GitHub: @BangerSoul
- Email:
contact@bangersoul.dev(or via GitHub profile discussions) - Timezone Overlap: Flexible overlap with US (PST/EST) and Europe (CET/GMT)
- Communication: Async-first, high-cadence PR deliverables with verifiable test proof


