AI Systems · Deep-Dive Writing · Vectr
Swapnanil Saha
I build AI systems and write long-form engineering deep dives. Creator of Vectr — semantic search and working memory for AI coding agents. Nine years of real-time distributed systems underneath.
Latest Deep Dive — 18 July 2026
Embedding Dilution: Why Semantic Code Search Misses the Answer
The docstring paraphrased the query — and still ranked below 200 look-alikes · 21 min read →
Deep-dive essays
13
Retrieval · Memory · Agents
Vectr
v1.3
Search + memory for AI agents · MCP
Compactions instrumented
246
Delivery-not-storage research
Years in production systems
9
Global ad exchange
Writing
Deep dives.
Measured, not summarized.
Long-form engineering essays on retrieval, agent memory, and context engineering — built on instrumented runs and honest ledgers, with every number traceable to an artifact.
17 July 2026 · 28 min · Claude Code
What Actually Survives /compact in Claude Code
108 + 138 forced compactions, graded token by token. What the boundary keeps, what silently dies, and where to put state so it survives.
23 May 2026 · 90 min · Foundations
The Complete Guide to Text Embeddings, Vector Databases & LLMs
Tokenization to HNSW graph search, with seven interactive demos — real tiktoken, real embeddings, a real 1,000-node index in your browser.
11 July 2026 · 22 min · Multi-Agent
The Seven-Agent Test That Gated Vectr 1.1
Seven AI agents build and review a real product coordinating only through shared working memory — the release gate unit tests can't be.
All 13 essays
swapnanilsaha.com/blog →
Vectr
Semantic search.
Working memory.
Open source · MCP · Zero config
Vectr gives AI coding agents the two things they lack: retrieval that understands code, and memory that survives.
An AST-aware codebase indexer with hybrid semantic + BM25 search, a symbol graph, and a persistent working-memory tier delivered through hooks — findings survive context compaction and session restarts instead of dying with the conversation. Any MCP-compatible AI code editor connects automatically. v1.1 added team mode — many agents, one shared memory; v1.3 adds per-memory triggers, so a note resurfaces exactly when its path, symbol, or lifecycle condition matches.
About
Systems depth.
Business clarity.
These days I build Vectr and write long-form engineering deep dives — that is what this site leads with. Underneath it sit 9 years inside the engine room of real-time digital advertising, building infrastructure that decides, in milliseconds, which ad a user sees: a custom Redis module that cut filtering latency 5×, auction strategy systems that directly moved advertiser budgets, real-time budget pacing on Cassandra across billions of daily requests.
What makes my work unusual is that I've never been able to hide behind pure engineering. Ad-tech has immediate, measurable outcomes — every latency improvement has a win-rate implication, every auction strategy change has a revenue footprint. Connecting technical decisions to business results is a discipline I've been building for a decade.
I also quietly became the person who shipped internal AI tooling before it was fashionable — a SERP template generator, a Slack-integrated docs wiki, and a campaign agent that can pause underperforming campaigns autonomously. Each one started from a business problem, not a technology impulse. That's the pattern I intend to keep.
Current role
Lead SDE · Global ad exchange
Location
Mumbai, India
Primary domain
Ad-Tech · Real-Time Systems
Building
Vectr · The deep-dive blog
Education
B.E. CS · BIT Mesra
Founder instinct
2 startups built · 2015–16
llm-tools suite
Six tools.
Shipped.
Six production-grade tools — each solving a real enterprise problem with LLM APIs. Built independently, in public, to the same standard I'd apply to internal tooling at scale. Each ships with Docker, a REST API, input validation, and a full landing page.
Developer Tools · Code AI · MCP
vectr
Zero-config codebase indexer for AI assistants. AST-aware chunking, hybrid semantic + BM25 search, and cross-session memory via MCP. Your AI code editor — Claude Code, Cursor, or any MCP client — gets only the relevant chunks, not entire files. Benchmarked on Apache Camel: −58% Phase 2 cost, −72% tool calls.
Compliance · DPDP Act · India
dpdp-copilot
Self-hosted operator tool for India's DPDP Act. Public intake form accepts data principal requests in 5 Indian languages. Automatically classifies and drafts a compliant multilingual reply. Operator inbox shows live SLA status. Every action logged as immutable evidence — exportable as PDF or CSV for regulatory audit.
AI Architecture · RAG
rag-readiness
Opinionated architecture recommendation per component. Diagnoses existing broken stacks with root causes by severity. Rule-based cost estimation, iterative refinement with session persistence, and RAGAS eval dataset generation.
LLM · Evaluation · Regression · CI/CD
llm-eval-suite
Evidence-backed multi-dimensional scoring across 10 task types including RAGAS-compatible RAG eval. Regression testing saves baselines and blocks CI when scores drop. Claim-level hallucination detection, prompt sensitivity analysis across variants, and multi-judge panel consensus with disagreement flagging.
Enterprise AI · Strategy · LangGraph
ai-use-case-scoper
Hybrid 8-question flow with optional Graph RAG doc enrichment — pre-fills tech context, always asks qualitative questions. LangGraph check-in pipeline evolves plans. Company memory persists across sessions.
Productivity · Meeting AI
meeting-to-action
Normalizes Zoom/Teams/Meet transcripts, detects meeting type, extracts structured decisions and action items. Cross-meeting commitment tracker flags missed items by severity. Weekly digest aggregates open work by owner across all sessions.
Skills
What I
bring.
Systems · Infrastructure
Distributed Systems Design
Redis · Custom Modules (C)
Apache Cassandra
Real-Time Data Pipelines
High-Throughput Ad Serving
Observability · SRE
AI · LLM Deployment
LLM Application Development
RAG Pipelines
LangChain · LangGraph
MCP Server Development
Eval Frameworks (RAGAS)
Agentic Workflows · Actions
Domain · Business
Real-Time Bidding (RTB)
Auction Mechanisms (1st/2nd)
Ad Serving Architecture
CPA / Conversion Tracking
Python · JavaScript · C
Founder Mindset
Experience
Nine years.
One company.
All nine years at one global ad exchange — specifics on LinkedIn.
Real-time financial infrastructure + three internal AI tools shipped
Two parallel tracks at the Lead level — one solving the hardest distributed consistency problem in ad-tech, one establishing the company's LLM practice from scratch.
Business Problem
Ad campaigns must pace spend accurately across thousands of real-time auctions per second. Underspend wastes advertiser budget; overspend creates financial liability. In a distributed system, maintaining a consistent, real-time view of "how much has been spent" while sustaining the write throughput that RTB demands is a fundamentally hard problem — and getting it wrong is immediately visible in the P&L.
Engineering Solution
Architected real-time budget pacing infrastructure on Apache Cassandra — engineered for financial precision under distributed consistency constraints. Designed the state synchronisation layer that keeps budget counters accurate across geographically distributed nodes without sacrificing the write throughput that real-time bidding demands. Correctness is non-negotiable; the system either works exactly right, or advertisers notice immediately.
On the AI front, shipped three internal tools — each starting from an identified business pain:
Auto SERP Template Generator — eliminated hours of manual ad template creation. Became the most-used AI tool companywide. The first proof that LLM deployment inside the company was worth investing in.
Docs Wiki · Slack Integration — made internal knowledge accessible via natural language through Slack webhooks. Reduced time-to-answer on process and technical questions that previously required pinging specific people.
Campaign QnA Agent — a conversational assistant that goes beyond answering questions. It can take minor actions autonomously — including pausing underperforming campaigns — making it the first agentic tool in the company's stack.
Auto SERP Template Generator — eliminated hours of manual ad template creation. Became the most-used AI tool companywide. The first proof that LLM deployment inside the company was worth investing in.
Docs Wiki · Slack Integration — made internal knowledge accessible via natural language through Slack webhooks. Reduced time-to-answer on process and technical questions that previously required pinging specific people.
Campaign QnA Agent — a conversational assistant that goes beyond answering questions. It can take minor actions autonomously — including pausing underperforming campaigns — making it the first agentic tool in the company's stack.
Auction strategy systems — engineering market mechanisms that move money
Business Problem
The digital advertising market was mid-transition from second-price to first-price auctions — a fundamental market-structure shift changing how advertisers bid and how publishers yield. The exchange needed to correctly implement both mechanisms and ship bid shading logic that protected advertiser ROI while preserving publisher yield. Get the balance wrong and advertiser budgets migrate to competing exchanges within days.
Engineering Solution
Designed and shipped first-price and second-price auction engines plus competitive pricing and bid shading logic — navigating the inherent tension between supply-side and demand-side incentives in a real-time marketplace. Built simulation tooling to evaluate pricing strategy changes pre-deployment, enabling data-driven decisions with quantifiable revenue impact before touching production traffic.
Auction strategy changes have direct, immediate revenue footprints — this is engineering where every decision is visible in the P&L within hours. Operated at the intersection of market economics, systems engineering, and business strategy; translated partner feedback and market signals into algorithm changes with measurable outcome metrics.
Conversion tracking SDK — developer experience as a direct lever on business outcomes
Business Problem
Advertisers couldn't reliably measure whether users who clicked their ads went on to convert — buy, sign up, complete a form. The existing pixel was brittle and frequently misimplemented, producing noisy conversion signals that made CPA bidding unreliable. Without clean conversion data, the entire bidding optimisation stack is operating on bad inputs. This was limiting budget growth from performance advertisers who need to prove ROI.
Engineering Solution
Redesigned the conversion tracking pixel JavaScript SDK from scratch — prioritising correct implementation by default and minimal integration friction. Achieved rapid, complete adoption across the full advertiser base. Accurate CPA measurement fed downstream bidding optimisation, producing measurable improvement in cost-per-acquisition efficiency and directly increasing advertiser budget commitment on the platform.
Owned the full lifecycle of a customer-facing technical artifact: requirement scoping → implementation → cross-functional coordination with business and account teams → rollout → post-launch measurement. The same end-to-end ownership pattern that defines every successful enterprise AI deployment — just an earlier iteration of it.
The Redis module — 200ms → 40ms, written in C, independently
Independently engineered a custom Redis module in C that reduced ad filtering response latency from ~200ms to ~40ms — a 5× improvement — while sustaining significant QPS gains. This directly enabled tighter real-time bidding windows and improved win rates for downstream auction systems. Also owned reliability and observability of the serving infrastructure handling billions of daily ad requests, maintaining sub-SLA latency targets across distributed components and transitioning the team from reactive incident response to proactive system ownership.
Earlier · Startup Experience · 2015–2016
India's most downloaded shopping extension — early-stage product engineering
Contributed to Flipshope during its early growth phase — a browser extension that solves a genuine consumer problem: detecting real discounts versus manufactured price inflation, and automatically applying the best coupon at checkout. The product went on to become the most downloaded shopping Chrome extension in India, with 5L+ extension installs, 10L+ app downloads, and a 4.5+ rating on the Chrome Web Store across 300+ supported brands. Worked in a bootstrapped, founder-led team where shipping fast and shipping right were the same thing.
First venture — zero to live product as a student
Founded and built Ransphire while still in college — going from idea to a live product entirely independently. The experience of identifying a problem, building a solution without a playbook, and putting it in front of real users planted a business instinct that has quietly shaped how I approach engineering problems ever since. The company didn't scale. The education was irreplaceable — and it's part of why I think about technical work the way I do.
Now
What I'm
building.
Active — Vectr
Shipping the memory half of Vectr
v1.3 shipped a per-memory trigger engine — path, symbol, semantic, and lifecycle conditions per note — on top of 1.1's team mode and 1.2's opt-in proactive injection. Current focus: the delivery-not-storage research line — a controlled study of agent memory under repeated context compaction (108 and 138 forced boundaries, token-level grading) — and an upstream feature request to Claude Code (#78795) proposing triggered injection for auto-memory files, so the findings become platform behavior rather than a workaround.
Writing — next up
Three deep dives in the pipeline
Embedding dilution (why code search misses the function whose docstring answers the query), deleting every ranking heuristic from a search engine, and why models won't call your MCP tool — each built on measured runs, like everything else here.
Retrieval research
Understanding-grade code retrieval
Moving Vectr's search from embeddings-plus-heuristics to structural signals: reference-graph importance priors and per-symbol purpose vectors, evidence-gated on public corpora before anything ships.
Contact
Let's
talk.
If you're working at the intersection of production systems and enterprise AI — or if something here sparked a conversation worth having — I'm always open to a good discussion.
Nine years in ad-tech means nine years of technical decisions with same-day revenue consequences. That feedback loop is hard to replicate — and it changes how you build.