Road safety systems built at SaveLIFE Foundation, before MetaMinds
Three production systems: a multi-stage retrieval pipeline, a WhatsApp learning channel built for exactly-once delivery, and a computer vision pass over dashcam footage. All three were built by Aniruddh Atrey at SaveLIFE Foundation, a prior employer. MetaMinds was not the contracting party, and that goes here rather than in a footnote.
Who the client was, and it was not MetaMinds.
This changes how every number below should be read, so it goes before them.
The three systems on this page were built by Aniruddh Atrey at SaveLIFE Foundation, a road safety organization, as work he did while employed there. MetaMinds was not the contracting party. There was no statement of work between MetaMinds and SaveLIFE Foundation, no invoice, and no client relationship of any kind. SaveLIFE Foundation is a prior employer of our founder, not a customer of this firm, and nothing here should be read as an endorsement by them or as work they commissioned from us.
It is published anyway, with the caveat first, for a reason that serves your evaluation as much as ours. You are trying to answer a specific question: what class of problem has this firm actually solved, and by what mechanism? A vague client story, a leading NGO, a national safety program, answers none of that, and the vagueness is doing work: it hides the one thing you would want to check. Naming the organization and stating plainly that it was not our engagement gives you something falsifiable instead.
Three systems, three surfaces, one scarce resource.
Retrieval, exactly-once messaging and computer vision share almost no components. What they share is the economics: in each case work arrived faster than a qualified person could review it, and hiring more reviewers was not the available answer.
Retrieval
Multi-stage Haystack v2 with e5-base-v2 embeddings over three MongoDB Atlas vector and text index pairs.
Analyst lookup time reduced approximately 70%, across 100+ reports per week.
Measured at SaveLIFE Foundation, a prior employer, not a MetaMinds client.
Delivery
A WhatsApp learning channel on a 12-state machine, with HMAC-SHA256 timing-safe webhook verification, Redis SET NX idempotency, a dual LRU and Redis cache, and a circuit breaker.
114 videos, 18 assessments and 99 learning journeys, at approximately 70 messages per second.
Measured at SaveLIFE Foundation, a prior employer, not a MetaMinds client.
Vision
A computer vision pass over dashcam footage, surfacing candidates for human review rather than replacing it.
500+ hours of footage at 95% precision across 200+ km of national highway. Manual audit effort reduced 85%.
Measured at SaveLIFE Foundation, a prior employer, not a MetaMinds client.
How was the road safety research corpus made answerable in seconds?
At SaveLIFE Foundation, a prior employer of our founder and not a MetaMinds client, analyst lookup time was reduced approximately 70%, across 100+ reports per week. That is the measured result, and it came from a multi-stage retrieval pipeline built on Haystack v2, with e5-base-v2 embeddings, over three MongoDB Atlas vector and text index pairs.
The detail worth pausing on is the pairing. Each of the three index pairs is a vector index and a text index over the same content, not one or the other. That is the same choice CourtNetra makes, and it exists because dense and lexical retrieval fail in opposite directions. An embedding is good at paraphrase and bad at exact tokens: a section number, a road identifier, a specific report title, a proper noun. Those are exactly the strings a domain analyst types, and exactly the strings an embedding smooths into its neighbors. A text index does not smooth them. Running both over the same content and reconciling the results costs more than running one, and it is still cheaper than trying to make either one do the other’s job.
Multi-stage matters for the reason a single retrieval pass is a bad default anywhere. One pass commits to whatever the first ranking returned, with no opportunity to notice that the retrieved set does not actually address the question. Stages give the pipeline somewhere to make that observation and act on it before generation, rather than generating fluently over a weak context and hoping the reader catches it.
How do you deliver training over WhatsApp without dropping a message or replaying one?
At SaveLIFE Foundation, a prior employer and not a MetaMinds client, the delivery system Aniruddh Atrey built carried 114 videos, 18 assessments and 99 learning journeys, and ran at approximately 70 messages per second. The engineering underneath it is almost entirely about one property messaging platforms do not give you: each learner receives each step once, in order, though the transport guarantees neither.
- A 12-state finite state machine
- Holds conversation position. The alternative most systems reach for first is a bag of session flags updated wherever the code happens to touch them, which works until two messages from the same learner arrive close together and leave the session in a combination nobody designed. An explicit state machine makes the legal transitions enumerable, which means the illegal ones are detectable rather than merely unlikely.
- HMAC-SHA256 verification, with timing-safe comparison
- Authenticates inbound webhooks. The signature check is the obvious part. The timing-safe comparison is the part that gets skipped: a naive byte-by-byte comparison returns early on the first mismatch, and the time it takes to return is a measurable function of how many leading bytes were correct. That is a side channel, it is exploitable given enough samples, and the fix costs one function call.
- Redis SET NX for idempotency
- Messaging platforms retry, which means the same inbound event arrives more than once, and a duplicate that advances a learning journey by one step is a real defect visible to a real person. SET NX is an atomic claim on a message identifier: the first processor to set the key owns the message, and every later copy finds the key present and stops. Not a check followed by a write, which has a race between the two, but one atomic operation.
- A dual LRU and Redis cache, behind a circuit breaker
- The in-process LRU absorbs the hottest reads without a network hop; Redis is the shared tier behind it. The circuit breaker decides what happens when a dependency degrades: it stops the system converting one slow downstream into a queue of stalled requests, and it makes the degraded behavior a designed state rather than an emergent one.
How do you audit 200+ km of highway without watching 500+ hours of video?
At SaveLIFE Foundation, a prior employer and not a MetaMinds client, a computer vision pipeline Aniruddh Atrey built processed 500+ hours of dashcam footage at 95% precision across 200+ km of national highway, and reduced manual audit effort 85%.
The economics are the whole argument. Reviewing dashcam footage is linear in the length of the footage and requires a trained reviewer for every minute of it. Over 500 hours of video is over 500 hours of review time. A vision pass that surfaces candidates turns a linear watch into a review queue. What is verified is the 85% reduction in effort, not the review workflow behind it, so this page does not describe one.
Read the 95% carefully, because it is precision and not accuracy and not recall. High precision with unmeasured recall produces a queue in which almost everything is worth looking at, which is exactly what you want for reviewer efficiency, and it produces no evidence at all about what went unseen. Whether that trade is acceptable depends entirely on the consequence of a miss, and that is a decision for the domain, not for the model.
Five limits, stated before you find them.
Each pairing below is a real gap. Naming them here is the same discipline MetaMinds sells, applied to work MetaMinds did not sell.
A lookup-time reduction is not an accuracy measurement
An approximately 70% reduction in lookup time says how long the work took, not whether the answers were right. Those are different measurements and only one of them was made. There is no published grounding rate for this system. The under-2% CI gate elsewhere on this site is a property of CourtNetra, measured against CourtNetra's own evaluation set, and it does not transfer here by association.
Throughput is not engagement
Approximately 70 messages per second is a capacity figure and 114 videos is an inventory figure. Neither tells you how many people completed a journey or what they learned, and we hold no verified figure for either, so there is none on this page.
95% is precision, not accuracy and not recall
Precision at 95% means that of the detections the system flagged, about 95 in 100 were real. It says nothing whatsoever about how many real instances went unflagged. That number is recall, recall is a different measurement, and we do not hold one, so it is not here. A vendor quoting a single accuracy percentage for a detection system is either eliding this distinction or has not thought about it, and it is the first question to ask them.
No safety outcome is claimed
Nothing on this page says these systems reduced crashes or saved lives. No such figure exists in our claims register and we will not manufacture one to make an engineering page more moving than it deserves to be. What is verified is what the systems did to the work of the people operating them.
The detection classes are not published
What the vision model was trained to find is not something we publish, so the pipeline and its measured properties are described and the rest is left out rather than filled in with a plausible-sounding list.
What carries across to a system that is not about road safety.
The expert is the constraint, not the model
All three systems address the same economics: work arriving faster than a qualified person can review it. That is the same shape as a claims queue, a contract review pile or a diligence data room, and it is the condition retrieval, automation and detection are built for.
Hybrid retrieval, for the same reason both times
Each of the three index pairs is a vector index and a text index over the same content. Dense and lexical retrieval fail in opposite directions: an embedding is good at paraphrase and bad at exact tokens, and exact tokens are what a domain expert types. CourtNetra makes the same choice, reconciling dense pgvector similarity and tsvector BM25 by reciprocal rank fusion.
Exactly-once is engineering, not configuration
A messaging platform gives you at-least-once and out-of-order. Every property above that is something you build: an explicit state machine, an atomic idempotency claim, a signature check that is not itself a side channel. None of it is exotic and all of it is skipped under deadline.
State the metric you did not measure
Precision without recall, throughput without engagement, speed without accuracy. Each pairing above is a real gap, named on the page rather than left for a buyer's own analyst to find. That habit is what MetaMinds sells; this page is it applied to work MetaMinds did not sell.
Asked and answered.
- Was SaveLIFE Foundation a MetaMinds client?
- No. The systems on this page were built by Aniruddh Atrey while employed at SaveLIFE Foundation, before MetaMinds existed. There was no statement of work between MetaMinds and SaveLIFE Foundation, no invoice, and no client relationship. SaveLIFE Foundation is a prior employer of the founder, not a customer of this firm, and nothing here should be read as an endorsement by them.
- Why publish prior-employer work as a case study at all?
- Because the question a buyer is actually asking is what class of problem this engineer has solved and by what mechanism, and a vague client story answers none of that. Naming the organization, naming the engineer and stating plainly that it was not a MetaMinds engagement gives something falsifiable instead. MetaMinds has one case study of its own, CourtNetra. Dressing prior work as a client engagement would inflate that by inventing a commercial relationship that does not exist.
- What does the approximately 70% lookup time reduction actually measure?
- At SaveLIFE Foundation, a prior employer of our founder and not a MetaMinds client, it measures how long it took an analyst to find a specific finding in the corpus, before and after the retrieval pipeline was introduced, across more than 100 reports a week. It is a measurement of time spent, not of correctness. There is no published grounding or hallucination rate for this system, and the under-2% CI gate quoted elsewhere on this site belongs to CourtNetra and does not transfer to it.
- What does 95% precision mean for the dashcam vision system?
- The figure is from a computer vision pipeline built at SaveLIFE Foundation, a prior employer and not a MetaMinds client. Precision at 95% means that of the detections the system flagged, approximately 95 in 100 were real. It says nothing about how many real instances the system failed to flag, which is recall. Recall was a separate measurement and is not one we publish. High precision with unmeasured recall gives a review queue in which almost everything is worth looking at, which is good for reviewer efficiency and produces no evidence about what went unseen.
- Why does a WhatsApp delivery system need a state machine and idempotency?
- The system in question was built at SaveLIFE Foundation, a prior employer, not for a MetaMinds client. It needs both because the transport guarantees neither exactly-once delivery nor ordering. Messaging platforms retry, so the same inbound event arrives more than once, and a duplicate that advances a learner one step is a defect a real person sees. A 12-state machine makes conversation position explicit so illegal transitions are detectable, and a Redis SET NX key is an atomic claim on a message identifier so the first processor owns it and later copies stop.
- Why is timing-safe comparison worth mentioning on a webhook handler?
- Because a naive byte-by-byte comparison of an HMAC signature returns early on the first mismatch, and how long it takes to return is a measurable function of how many leading bytes were correct. Given enough samples that is an exploitable side channel. The fix is a single constant-time comparison function, which is why its absence signals a handler built quickly rather than carefully.
- Do these numbers predict what MetaMinds would achieve on my system?
- No. Every figure on this page is past tense and attached to a named system at a named organization. A percentage measured on a road safety corpus with its own baseline, its own analysts and its own workflow is evidence that the class of problem is solvable and it is not a forecast for a different corpus. Anyone quoting you a number before seeing your workload is reading it off someone else's system.
- What was chosen over this case study, and why?
- Two alternatives were considered. The VelaReach microservices work, 9 Dockerized services with 122 API endpoints, 807 tests and a Kafka event backbone, is real and substantial but is platform engineering rather than an AI system, so it demonstrates less of what MetaMinds sells. The INNEFU Labs agentic triage work over Splunk SIEM with LangGraph, which took SLA attainment to approximately 90% and analyst throughput to 3x, is closer to the offer but has too little publishable depth to sustain a full technical narrative. The SaveLIFE Foundation work won because three separate verified systems can be described in mechanism rather than summary.
The system MetaMinds actually owns.
CourtNetra is ours, it is live, and every figure on it is a production number: hybrid retrieval over 18,863,754 judgments, with a weekly CI evaluation that fails the build above 2%.
Every figure on this page belongs to a prior employer. Every figure on that one belongs to MetaMinds.