#01The origin
issue
Software engineer ✦ AI systems ✦ Backend Vol. — India

Ratnam Ojha

NarrationI’m Ratnam Ojha. I design clear, resilient systems for the messy problems that matter.

Scroll to begin Read the origin

Chapter 01Origin

A builder with a systems mindset.

Narration

I’m drawn to the place where useful software meets difficult infrastructure: agentic systems, backend platforms, and the small details that make products feel inevitable.

My approach blends careful engineering with product instinct.

Powers

  • A

    AI / ML systems

    Agentic workflows, retrieval, and reasoning layers.

  • B

    Backend engineering

    Platforms that stay calm under pressure.

  • C

    Product instinct

    The details that make software feel inevitable.

Base of operations

India · working globally

28.6139° N / 77.2090° E

Trace the thread. Understand the constraint. Make the next move count.

— Operating principle

Chapter 02Missions

Where I’ve built.

Selected roles and high-ownership work across product, AI, and infrastructure.

  1. June 2026 – Sept. 2026San Francisco · RemoteComplete

    Software Engineer Intern

    Omi AI

    • Isolated a 50-minute total transcription outage across application, provider API and network layers, then rebuilt the provider fallback chain so the path degrades instead of failing PR #11814
    • Redesigned public developer API upload limits from one global ceiling to per-route limits, closing a rate-limit gap that had doubled per-key throughput PR #8251
    • Found and fixed a prompt-injection route that let private user data leave through a server-side tool PR #11857
    • Closed four silent failure points where a provider refusal was indistinguishable from silence PR #11090
  2. July 2026 – Aug. 2026Gurugram · HybridComplete

    Software Development Engineer Intern

    Dcluttr

    • TypeScript
    • Mastra
    • MCP
    • PostgreSQL
    • ClickHouse
    • Fixed two anti-hallucination verifier bugs that marked valid figures untraceable; added prior-turn and aggregate evidence checks, then confirmed fabricated figures were still rejected on live warehouse data
    • Cut worst-case verifier latency from 1.9s to 16ms by replacing O(n²) pairwise search with sorted-window binary search
    • Built a 200K-case differential fuzzing harness against the original implementation, catching a regression that 27 passing unit tests missed
    • Reduced stored conversation size 70% (1.82MB to 553KB over 302 messages) by dropping tool traces before persistence, saving ~318K estimated recall tokens
  3. Aug. 2025 – Dec. 2025New DelhiComplete

    Research Intern

    IIM Ranchi

    • SMOTE
    • BERT
    • Identified class imbalance in educational chatbot datasets as a key barrier to adoption, leading to biased intent recognition and poor student experience
    • Applied SMOTE-based resampling and distribution analysis to balance minority intent classes, improving model fairness and robustness across underrepresented student queries
    • Evaluated transformer-based architectures (BERT) to better capture contextual and semantic nuances, enabling more reliable and adaptive conversational responses for diverse learning needs
  4. July 2025 – Aug. 2025Complete

    AI Engineer Intern

    Kaiten Software

    • GPT-4.0 API
    • n8n
    • ERPNext
    • Whatsera
    • Addressed high manual customer support overhead by developing AI-powered WhatsApp and website chatbots using GPT-4.0 APIs, significantly automating first-level customer interactions
    • Designed and deployed an end-to-end automated lead management pipeline using n8n workflows and ERPNext, enabling AI agents to autonomously collect, track, and follow up on leads until human intervention was required
    • Orchestrated multi-platform system integration across ERPNext (CRM), n8n (workflow automation), and external APIs (Whatsera, GPT-4.0), resulting in a cohesive, scalable customer engagement and lead-nurturing ecosystem

Chapter 03Arsenal

Things that connect.

A few gadgets from the workbench. Each one begins with a precise problem, then earns its complexity.

01 / AI agents × payments

IntentGuard

A policy-enforcement gateway between autonomous AI agents and payment systems. Every action gets allow, human review or deny — from signed intent, spend caps and risk scoring.

p95 latency
9.6ms
Requests / sec
~2,000
Overspend, 8 replicas
0

Python · FastAPI · PostgreSQL · OPA/Rego · Redis · Next.js · Docker

02 / Agri-data × ML

Cerealia

Ranks crops across 36 Indian states and UTs by risk-adjusted return, not one predicted crop — an ML suitability model combined with yield, price, water and risk data.

Govt. records
246K
Districts
646
Crops
124

Python · FastAPI · scikit-learn · pandas · React · D3 · Groq

03 / MCP × Search

ArXiv Research Intelligence Server

An MCP server exposing a live paper index to language models: six tools over search and document QA, an edge-case suite, and a deploy on every push.

MCP tools
6
Index per doc, from 8s
<1s
Replaced 90MB embeddings
BM25

Python · MCP · BM25 + RRF · asyncio · GitHub Actions

Chapter 04Beyond the mask

The rest of the story.

Good engineering comes from paying attention. Mine is shaped by cinema, open internet rabbit holes, and the occasional late-night prototype.

Allies & hideouts

Chapter 05The signal

Send the signal.

Have an interesting problem? Light it up — I’ll answer.