Senior Software Engineer at New Relic. Seven years on distributed systems, real-time CDC pipelines and fault-tolerant services — at Oracle, Goldman Sachs, Paytm and FICO. The AI side is what I am building toward, in the open, on the single VPS that runs a six-agent system.
Now · Nov 2026
Rebuilt this site around the ledger. Next is the eval harness the experiments will run on — it does not exist yet — and then the first measured row, HNSW recall under index churn. At work, an SLI/SLO evaluation service over a Kafka stream.
Written 1 Nov 2026. If this line is stale, assume the rest is too.
01 / Findings
A vector index sitting at recall@10 = 0.98 can still ship a search system that feels broken, because recall@k measures fidelity to brute force over your own embedding space, not whether the right answer is in that space at all.
Preliminary Argued 12 May 2026 ↗
Past roughly 50 million vectors the HNSW index stops being the hard part. Shard semantically and the shard router becomes a second layer of approximate search: it fails silently, and it accounts for most of the recall loss.
Preliminary Argued 5 May 2026 ↗
02 / About
Seven years shipping production systems at scale: payment flows through NPCI, account lifecycle at a bank, real-time change capture out of a production database. The through-line is scale, resilience and fault tolerance — circuit breakers, replay, and the details that separate a demo from something that survives someone else's outage.
The AI work is newer and it is mine, not my employer's: a six-agent setup with hybrid Redis and pgvector memory, Telegram bots and a voice pipeline, running 24/7 on a single box, and writing down what it actually costs to run. I am learning this part in public rather than claiming it.
03 / Experience
New Relic · Bengaluru
An SLI/SLO evaluation service over a Kafka stream, where a periodic scheduler and an event consumer have to agree about the same minute.
Oracle · Bengaluru
Real-time CDC streaming with GoldenGate and a WebSocket event layer, and a DAG orchestration engine with cycle detection and stateful failure recovery.
Goldman Sachs · Bengaluru
A metadata-driven interface for account lifecycle events, and the Grafana and Prometheus stack that made trading-platform reconciliation debuggable.
Goldman Sachs · Bengaluru
Post-account-opening workflows, built from scratch.
Paytm · Noida
FasTag issuance and payment workflows through NPCI, plus a centralised in-memory config store and the circuit breakers that stopped one slow dependency taking the rest down.
FICO · Bangalore
Backend services for customer onboarding and loan repayment.
NIT Bhopal
Spectrum hole detection for cognitive radio, and a paper on ensemble methods for fraud detection on heavily skewed data.
04 / Build
Six specialised agents on one VPS: hybrid Redis and pgvector memory, a Telegram front end, a voice pipeline, and scheduled passes that run without me.
CDC pipelines with GoldenGate; real-time replication, and the coordination problems that appear when two clocks disagree.
Terraform, OCI, distributed system design, and production deployments that have to survive someone else's outage.
CLI tools in Go, a local text-to-speech pipeline, and LLM inference tuned for machines without a GPU.
05 / Stack
06 / Writing
Graph rot under writes and the hot/cold index pattern that answers it, redundancy in pure similarity ranking, and why recall@k measures your approximation rather than your product.
The Dev Kernel · Part 2 of 2
What changes past tens of millions of vectors: semantic sharding makes the index tractable and creates a router, and that router is a second layer of approximate search that fails silently.
The Dev Kernel · Part 1 of 2
07 / Contact
Happiest talking about stream coordination, agent evaluation, or why your index is quietly rotting.
ankitmishra471@gmail.com