Skip to content
View allands365-hub's full-sized avatar
  • mumbai
  • 18:49 (UTC +05:30)

Block or report allands365-hub

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
allands365-hub/README.md

Hey, I'm Allan D'Souza

AI Engineer based in Mumbai — agentic systems, retrieval, and memory.

I build agents that run real workflows: multi-step orchestration with human approval gates, knowledge and memory layers agents can actually rely on, and the boring-but-essential parts — evals, budgets, fallbacks — that make them safe to point at a business.


What I Do

  • Agent orchestration — stateful multi-step workflows, tool/function calling, human-in-the-loop approval gates, webhook and schedule triggers, bounded loops with latency budgets and graceful degradation
  • Retrieval & memory — chunking, embeddings, vector search, reranking, knowledge graphs, citation verification and anti-hallucination constraints
  • Wiring agents into real systems — email, chat, project management, observability, via APIs, webhooks, and MCP
  • Making it measurable — eval harnesses, population-level tests, contract-seam guards, token budgeting and cost tracking

Tech I Work With

Agents      multi-step workflows · tool calling · HITL approval · MCP · evals · model routing
Retrieval   embeddings (Voyage AI · FastEmbed) · pgvector · ChromaDB · reranking · Cognee/Kuzu
Backend     Python/FastAPI · Java 17/Spring Boot · TypeScript/Node · Postgres · SQL Server · Convex
Frontend    React 19 · Next.js 16 · TypeScript · Tailwind · Bun · React Native · Electron
Infra       Vercel · Railway · Cloudflare · Blaxel · Docker · GitHub Actions · Grafana/Prometheus/Loki

Featured Projects

Agentic systems

Project What it does Tech
Distill AI bug-triage & release operator — multi-step workflow with a human approval gate, wired into Gmail, Slack and Linear. Webhook auto-triage on inbound mail. Lemma SDK, Python, agent workflows
MindMeld Every meeting folds into one queryable knowledge graph — full memory lifecycle: remember, recall across meetings, enrich, forget Next.js 16, FastAPI, Cognee, Groq Whisper, Kuzu
KSP Crime Intelligence Crime intelligence platform over 674k records — geospatial hotspot detection, offender network analysis, forecasting & anomaly detection, plus a bounded NL Ask-AI agent (typed tool registry, schemas derived from signatures, live eval at 8/10 tool selection) running fully on Zoho Catalyst Next.js, deck.gl, ECharts, FastAPI, scikit-learn, Zoho Catalyst (QuickML LLM)
OSS Warden AI-powered GitHub issue triage in ~10s — multi-agent, serverless CrewAI, MCP, Modal

Retrieval & applied ML

Project What it does Tech
MSME Health Card · live Explainable credit scoring for credit-invisible small businesses — WOE/logistic scorecard + ONNX challenger Next.js 16, TypeScript, optbinning, XGBoost→ONNX
Jaanch Lite Legal document intelligence — domain embeddings + instruction-following reranking + citation verification against 50+ Indian Acts Python, Voyage AI, ChromaDB, Instructor
Juristudio AI Canadian legal research — 28k cases, CREAC memos, 3-tier citation verification Next.js, Convex, Voyage AI
Coachsync Multi-tenant coaching-institute SaaS with semantic question matching Next.js, Bun, pgvector

Products & tools

Project What it does Tech
NoteSync Auto-detect meetings, record, transcribe & summarise — local-first, search across every past meeting Electron, Groq Whisper/LLaMA, SQLite FTS5
Bill Karo GST-compliant invoicing for Indian MSMEs — WhatsApp reminders, mobile-first React Native, Fastify, Neon, Drizzle
BuildAi Arena Hackathon platform connecting sponsors and builders Next.js, Supabase, Groq
AI-Newz AI newsletter generator — RSS to summarised email React, FastAPI, Groq
Agent Marketplace Discover and connect with AI agents Next.js, Supabase
PickMyTree Adopt a mango tree, follow it through the season — Razorpay checkout, webhook-driven fulfilment, farm-admin ops Next.js, Convex, Razorpay

Experience

Edmik / Indivue (Kademin) — AI Engineer · Feb 2026 – Present Own the AI layer of a production education platform (Spring Boot / SQL Server backend, four React frontends). AI feature-gating and permission system, token budgeting with cost-basis tracking, multi-provider model routing (Gemini 3 Flash → Kimi k2.5 fallback), and an LLM evaluation pipeline that grades handwritten student answers via OCR. Also deployed the Grafana MCP server on Blaxel so production metrics and logs are queryable over MCP. ~2,600 commits across 8 repos.

Allan Travels — Co-founder / Operator · 2017 – Present Car rental business. Built and still run the booking and billing automation it operates on.

Apar Games — Game Designer · 2012 – 2017 Shipped 10+ titles across Steam, iOS, Android and web — Scribbled Arena, Twilight Zone, Traffic Controller. Designed mechanics, feedback loops and in-game economies, iterating on live analytics.


How I Build Agents

Getting an agent to work once is the easy part. Making it behave predictably on the thousandth run, on input nobody anticipated, against a system where being wrong costs something — that's the job. Three things carry most of the weight:

Guardrails are part of the design, not a setting added later. Every agent I ship gets an explicit blast radius decided before the first prompt is written. Distill routes low-confidence and P0 items to human approval. The Edmik grader is advisory-only — accurate enough to be useful, not accurate enough to be trusted alone. The KSP agent is capped at two rounds and a 15s budget, and degrades gracefully instead of retrying into the void. The Grafana MCP server runs read-only, write tools disabled, on a Viewer-scoped token.

Knowing when not to build an agent. Distill's priority rubric and duplicate detection look like obvious LLM work and are deliberately plain Python, unit-tested — cheaper, repeatable, debuggable. A rubric that drifts between runs is worse than no rubric. The LLM does only what needs judgement.

Monitoring is the deliverable, not the follow-up. If an agent is reliable, there's an eval that says so with a number — KSP ships with a scripted live eval at 8/10 tool selection, 10/10 within budget. Tool schemas are derived from function signatures so they can't drift from the code, and a population test loops the full registry so a new tool can't skip coverage. LLM spend is tracked per feature and per tenant rather than discovered on an invoice. Production metrics and logs are queryable over MCP.


Let's Connect


Currently open to AI Engineer roles — agentic systems, retrieval, and applied LLM infrastructure.

Popular repositories Loading

  1. myportfolio myportfolio Public

    New portfolio

    JavaScript

  2. linkdnpost_gen linkdnpost_gen Public

    New landing page

    HTML

  3. AI-Newz AI-Newz Public

    TypeScript

  4. 25_hackathon 25_hackathon Public

    hackathon

    TypeScript

  5. Calculator_Test Calculator_Test Public

    testing for oss-warden

    Python

  6. allands365-hub allands365-hub Public