Python · Async · LiteLLM · Multi-Agent Swarms · Contextual RAG · DBOS · MCP
Open to Senior / Founding AI Engineer roles. Most recently the Founding AI Engineer / Systems Architect at Smart AI Technology Solutions (LawWorld) (Aug 2024 - Jun 2026).
I specialize in primitive-level orchestration. My work focuses on abandoning opaque multi-agent frameworks in favor of raw async Python, LiteLLM, and durable execution to enforce highly deterministic, transparent, and scalable ReAct loops with bounded iterations and strict schema validations.
"Frameworks obscure the cost ledger; primitives expose it."
1. Bare Agent | Framework-Free Multi-Agent Runtime
Published a zero-lock-in agent runtime to PyPI built on exactly 8 core primitives and zero dependencies. Designed for executing deterministic multi-agent chains with complete state transparency. Features a Next.js drag-and-drop studio that compiles visual node graphs directly into raw, high-performance async Python code (optimized for local Ollama execution at $0).
2. Argus | Multi-Agent Deep-Research Engine
A framework-free ReAct orchestration swarm built on Python, LiteLLM, PostgreSQL (pgvector), FastAPI, Kubernetes (KEDA), and DBOS.
- Cognitive Routing: Dynamically routes extraction tasks to fast models (Llama 3/Flash) and reserves frontier models (Claude 3.5 Sonnet) strictly for synthesis, yielding $0.149/turn unit economics.
- Reliability: Cohen's Kappa-calibrated LLM-as-a-judge eval gates over a golden dataset.
- Durable State: DBOS execution for crash-resumable state recovery on long-running research loops.
- Scale: KEDA scale-from-zero searcher pods on an ARQ Redis queue.
3. Agents Eval Framework | Autonomous Prompt Optimization
An automated prompt-calibration CI/CD suite built on DSPy and GEPA loops to enforce strict agreement thresholds prior to model regressions.
| Domain | Technologies |
|---|---|
| Languages | Python 3.12+ (Async, Pydantic, LiteLLM), TypeScript |
| Agentic Systems | MCP, Planner-Worker Swarms, Framework-Free Runtimes, Contextual RAG, Hybrid Search (RRF) |
| Evaluation | LLM-as-a-judge, Cohen's Kappa, DSPy, RAGAS, Golden-Set Replay |
| Infra & State | DBOS Durable Execution, Redis ARQ, Docker, Kubernetes + KEDA, Azure Container Apps, Terraform |
| Databases | PostgreSQL (pgvector), MongoDB Atlas |
| Observability | OpenTelemetry (OTel), structlog |
role: Founding / Sole AI Engineer (Smart AI Technology Solutions, Aug 2024 to Jun 2026)
status: Open to Senior / Founding AI Engineer roles
model_preferences: Claude 3.5 Sonnet (Synthesis), Llama 3 / Flash (Extraction)
focus: Agentic Systems, Applied LLMs, Production Backends, LLM Evals, Durable Execution
timezone: IST (UTC+5:30) | Async-First Remote
contact: subratasubha2@gmail.com