engineer: Abinesh U
role: AI Engineer · Agentic AI & LLM Systems
works_on: the layer between the model and production
principle: "Autonomy should be earned, and trust should be reversible."
status: open to AI Engineer roles · remote or relocationThe model is one node. Most agent failures happen in the rest of the graph: unclear decision authority, unmanaged state, no way to know why something happened. That is the part I build.
flowchart LR
REQ(["request"]) --> HAR
subgraph SYS["the system around the model"]
HAR["harness<br/>starts work, verifies, knows when to stop"] --> GRAPH["graph<br/>routes work across agents"]
GRAPH --> LLM(("LLM"))
LLM --> POL{"policy<br/>who decides?"}
POL -->|"confident"| ACT(["action"])
POL -->|"uncertain"| HUM(["human"])
ACT -.->|"loop: evidence"| HAR
STATE[("state<br/>what the system can trust")] -.- GRAPH
OBS["observability<br/>why it happened"] -.- POL
end
| Layer | Built in | Evidence |
|---|---|---|
| Decision authority | AegisAI: five parallel agents feed a policy engine; >30% disagreement escalates to a human; hash-chained audit log | 309 RPS · 0% errors · P95 500 ms at 200 concurrent users |
| Earned autonomy | Sollu: trust ladder where read-only tasks auto-run, soft tasks earn autonomy after 3 approvals, email never auto-runs | ~6.5 s median end-to-end on Cloud Run · grounded research 244 s → 18.7 s |
| Real-time orchestration | Hermes AI: Twilio → streaming STT → Gemini with agentic RAG → GPU TTS, with barge-in | Per-call TTFB and structured telemetry |
| Retrieval & routing | Athena: SLM-first RAG with Qwen2.5 query optimizer, BGE reranking, document RBAC, cost-aware inference routing | Circuit breaker · vLLM with fallback · OpenTelemetry tracing |
| Graph | agentic-patterns and Hierarchical-Agents: LangGraph patterns and a three-tier supervisor hierarchy | StateGraph · Command routing |
| Evaluation | Athena eval layer: golden set, recall@k, calibrated LLM-as-judge, CI regression gate | in progress |
Each piece takes one layer of the diagram above, at abinesh.blog.
| # | Article | Layer |
|---|---|---|
| 01 | Harness Engineering | harness |
| 02 | Loop Engineering | loop |
| 03 | State Engineering | state |
| 04 | Graph Engineering | graph |
| 05 | Agent Observability | observability |
Python TypeScript FastAPI · LangGraph Google ADK MCP LangChain · Gemini Qwen2.5 vLLM · Qdrant ChromaDB · PostgreSQL Redis · AWS GCP Cloud Run Terraform Docker Modal · OpenTelemetry Prometheus Grafana · PyTorch XGBoost SHAP MLflow · Deepgram Twilio
📧 abinesh@abineshu.dev · 💼 LinkedIn · 🌐 abinesh.blog · 🔬 ORCID