An agentic LLM pipeline built on LangGraph and MCP that turns unstructured WhatsApp payment messages into a 10-field Google Sheets ledger. It runs every 30 minutes, sends a daily summary, and guarantees exactly-once processing through SQLite deduplication, commit-after-write and locking, so no payment is ever logged twice.
94.8% → 100% held-out accuracy
92-message eval suite
Exactly-once ingestion
LangGraphMCPLLM EvalsSQLiteGoogle Sheets APIStreamlit
A multi-tenant workspace where LangGraph agents draft itineraries, marketing creatives, vendor deals and lead follow-ups, and a human approves every step before anything goes out. Tenant isolation is enforced in the database with PostgreSQL Row-Level Security, per-tenant integration tokens are encrypted at rest, and nothing is sent or posted except from the approve path.
Human-in-the-loop approvals
Row-Level Security tenancy
24 backend + 5 frontend tests
CI on every push
LangGraphFastAPIReactTypeScriptPostgreSQLpgvectorDockerTerraform
A support agent where the model proposes and ordinary code decides: no refund is ever issued without a human approval tied to one exact, immutable proposal. Refunds are idempotent and reconcile with the payment provider after lost responses. A bundled evaluation platform runs 50 versioned tasks against mock services with injected faults and scores business state, not wording.
No refund without approval
50 versioned eval tasks
4 deterministic evaluators
911 tests in CI
LangGraphFastAPIPostgreSQLpgvectorRAGGeminiLLM Evals
Modeled the public Delhivery dataset of 144,867 shipment records as a directed graph of 1,657 facilities and 2,783 corridors. I added node2vec and centrality features to a LightGBM ETA model and ranked every hub by SLA-breach and betweenness risk. Fixing just the top 3 bottlenecks projects 4% less late-delivery time.
MAE 107 → 32 min
−24% error vs non-graph baseline
5% → 46% legs within 15%
10× better than OSRM
NetworkXnode2vecLightGBMStreamlitGraph ML
Segmented 3,900 D2C customers into 5 loyalty tiers using MySQL CTEs, window functions and pandas features. Discounted customers spent no more than full-price ones (−$0.85 per customer), which exposed 43% of revenue ($99K of $233K) as promo-reliant. I turned that into a two-phase promo-sunset plan and a Power BI founder dashboard.
$99K promo-reliant revenue
281 Ideal Loyalists
680 at-risk customers flagged
SQLMySQLpandasPower BISegmentation
Three cooperating agents forecast demand, set a dynamic ₹/kWh tariff and learn from simulated outcomes. Built on real Caltech ACN-Data charging sessions and cross-checked on Shenzhen UrbanEV. Against a flat ₹15/kWh rate the learned policy earns more while easing peak congestion and moving demand into idle off-peak hours (simulated results).
+8.2% revenue vs flat rate
−22% peak wait
Load forecast R² 0.90
Congestion ROC-AUC 0.96
scikit-learnGradient BoostingSimulationPython
End-to-end SFT, reward model and token-level PPO on HelpSteer3 with Hugging Face TRL and LoRA, evaluated on 2,017 held-out prompts. PPO beats the base model in 50.9% and the SFT policy in 57.7% of pairwise comparisons under the learned reward model. The audit also shows its cost: longer outputs, more generation-cap hits and the highest repetition rate.
57.7% PPO win rate vs SFT
2,017-prompt evaluation
Higher repetition (audited)
TRLPPOLoRAQwen2.5PyTorch
