Mohit Aggarwal Software Developers Ltd

About

The company delivers enterprise software for financial services and payments. Engagements have included payment systems engineering for a global card network and portfolio and data tooling for a global asset manager, alongside data pipeline and platform work in agritech. The core stack is Java and Spring Boot, Kafka, AWS and Python.

Current work is in agentic AI systems: LLM applications with tool calling, retrieval, evaluation and human review, built to production standards rather than to demonstrate. The work spans the full path from research to running service, and the company operates what it builds.

Work

prior-auth-iq

Spring AI · PostgreSQL / pgvector · Java

A reference architecture for prior authorisation triage in healthcare claims. Requests are matched interface-first, so a deterministic rules path and an LLM path satisfy the same contract and can be swapped or compared. Determinations come back structured rather than as free text, with a human kept in the loop on anything the system will not decide alone.

Retrieval sits behind a configuration switch, running either in-memory or against pgvector with HNSW indexing, verified by a Testcontainers integration test that self-disables where Docker is absent. The wider bet is that AI arrives inside existing enterprise Java estates rather than replacing them.

github.com/aggarwamnew/prior-auth-iq

TransactGuard

Spring Boot · REST · Java

Anti-money-laundering transaction monitoring. Four detection rules covering structuring, velocity and threshold patterns, exposed over a REST API and covered by sixty tests. Built to mirror how monitoring actually runs inside a regulated institution, where an alert has to be explainable after the fact.

github.com/aggarwamnew/transact-guard

Dáil Monitor

Python · NLP · Dashboard

A natural language pipeline over Irish parliamentary debate, scoring more than twenty metrics across ten members and presenting them in a live dashboard. Covers the unglamorous half of applied NLP: acquisition, cleaning, and turning noisy transcripts into measures that survive scrutiny.

github.com/aggarwamnew/dail-monitor tcd-text-analysis.netlify.app

Way to Start

Flask · Firebase · Cloud Run · Stripe

An AI website builder, live with subscription customers. A description becomes a published site in seconds, refined through conversation, then deployed to a stable URL in one click. Streaming generation with real progress, a curated template gallery, and content moderation on the way in.

Runs on Cloud Run with Firebase authentication and Firestore persistence, behind 128 end-to-end tests that exercise real authentication, real deployment and real teardown rather than mocks.

waytostart.com

Chat Bridge

Flask · Gemini · Retrieval

A self-hosted assistant that injects a local knowledge base into each exchange, built to keep private context private while still using a hosted model. The precursor to the retrieval and memory work that followed.

github.com/aggarwamnew/chat-bridge github.com/aggarwamnew

Platform

The company runs its own self-hosted platform of seventeen always-on services, still growing, each one added when a need appeared rather than planned in advance. Operating it keeps deployment, monitoring and failure first-hand rather than theoretical. It spans three tiers.

Operational infrastructure

Reverse proxy with authentication, isolated service users and systemd unit management, scripted deployment, and layered backup to a second host and off-site. Centralised log aggregation across services and timers, with a REST API and a live streaming tail. Automated drift detection between source repositories and what is actually deployed. Scheduled network scanning: host discovery, port audit against a signed baseline, and host-level posture checks, with findings routed to alerting.

Intelligence

Daily gathering across tracked domains with density-based clustering, entity graphs and novelty detection, surfaced through dashboards. Semantic search over a private knowledge base combining embeddings, a keyword index, an entity index and a relationship graph, exposed to AI tooling over MCP.

Line-of-business tooling

Revenue and operations systems: reconciliation across tiered agreements, settlement and payout tracking, automated expense filing driven by browser automation against authenticated portals, and cash-flow modelling.

Research

Learning to Ask

Feedback-driven question generation · 2026

An outer learning layer treats a frozen language model as its environment and decides what to ask it next, guided by which questioning strategies have produced the most learning so far. Feedback-driven questioning produced knowledge graphs 4.9 times denser than random exploration.

A secondary result runs against intuition: structure-driven questioning without feedback performs worse than random, because the system settles into exploration ruts it has no way to detect.

doi.org/10.5281/zenodo.20113782 github.com/aggarwamnew/dream-layer

Contact

mohit@mohit.me github.com/aggarwamnew linkedin.com/company/mohit-aggarwal-software-developers-ltd