◉ TARGET//PRINCIPAL ENGINEER · APPLIED AI
MAOR BRIL
DISTINGUISHED ENGINEER · AI ENTERTAINMENT · CHARACTER.AI
I get handed the system that has to work when it is generating itself — and I build the thing that decides whether its output is good enough to ship.
Agentic systems in production at 20M+ monthly users, the Go SDK and post-trained models underneath them, and the evaluation wall that gates every release. Twenty years of infrastructure at Google, Netflix and Coinbase under all of it.
20M+
MAU ON MY AGENTS
AgentX: planner, executors, reviewers. 30+ agentic turns, no human in the loop.
~3s
THE RELEASE GATE
A trained 8B judge scores every generated clip before it ships. Open source.
8/12
DIRECT HITS ON THE ROLE SPEC
Scored line by line against a real posting. The other 4 say what I would be learning.
20+
ENGINEERS LED
Two orgs at Coinbase. Org-wide AI standards at Character.ai. Five systems on my SDK.
01 // THE PROOF
SHIPPED SYSTEMS // IN PRODUCTION
2/7 OPEN
// THE PIPELINE AND THE EVAL WALL LEAD. FOUR CONSUMER PRODUCTS FOLLOW AS PROOF THE SAME MACHINERY SURVIVES REAL USERS.
- PROMPTone request in
- PLANNERdecomposes the job
- EXECUTORSimage · video · audio
- REVIEWERStrained eval scores per stage
- RELEASE GATE~3s judge, in CI
- USERno human before this
The reliability problem of probabilistic systems, made concrete: every stage can produce something plausible and wrong, so every stage is reviewed by a different model with a different job before the next one runs.
Six specialized tools, bounded concurrency, full observability, and a release gate that has to clear before a user sees anything. Jeeves, an adaptive onboarding agent, runs on the same SDK with a different goal and a different tool registry — the platform is general, not one-off.
PLANNER / EXECUTOR / REVIEWER30+ AGENTIC TURNSTRAINED EVAL GATESBOUNDED CONCURRENCYGO · claude-agent-sdk-go20M+ MAU
MULTI-AGENT · TRAINED EVAL GATES · NO HUMAN IN THE LOOP
CLIP and LPIPS catch per-frame sharpness and caption match. They miss everything that decides whether a clip is usable: does it tell the story you meant, does it obey physics, does the character stay the same person, does the pacing land, is the audio synced.
JudgeJudy scores those axes per clip with a Qwen3-VL 8B model in about three seconds — small on purpose, because it lives at the gate and has to keep up with CI. We A/B'd a bigger model: marginally better, significantly slower, not worth it. The judge is not prompted, it is trained: pairwise preference — is B a better story than A? — fine-tuned with Bradley-Terry loss over real and deliberately corrupted footage, then calibrated against human annotations.
Judge Dredd is the layer above: it pulls Amplitude and Statsig telemetry over MCP, correlates offline scores against real engagement, and only the dimensions that predict behaviour become release gates. An eval nobody validates against users is a vanity metric.
PAIRWISE PREFERENCE TRAININGBRADLEY-TERRY LOSSHUMAN CALIBRATION LOOPQWEN3-VL 8B · ~3s / CLIPTELEMETRY-VALIDATED GATESOPEN SOURCE · MIT
TRAINED JUDGE · HUMAN CALIBRATION · TELEMETRY-VALIDATED GATES
CASE BOARD — FULLY PRODUCED MYSTERIES, GENERATED AND PLAYABLE IN SECONDS
EVIDENCE BOARD — AGENT-DRIVEN INTERROGATION, CONTRADICTIONS, TIMED ACCUSATION
Every hard problem in agent-run interactive fiction, in one product.
Agents generate a coherent mystery, hold character through interrogation, and keep world state consistent across evidence and testimony — while staying solvable, under a timer, in front of players who are actively trying to break it.
Every case is cast, filmed, scored and cleared by the eval wall before it reaches the board.
GENERATIVE NARRATIVEIN-CHARACTER AGENTSWORLD-STATE CONSISTENCYSOLVABILITY EVALSREAL-TIME PLAY LOOP
GENERATIVE NARRATIVE · IN-CHARACTER AGENTS · TIMED PLAY LOOP
PRODUCTION PIPELINE — SCRIPT GENERATION → AI CONTINUITY REVIEW → SEGMENT TTS → PUBLISH → PROMO VIDEO
The hard part is not generating one episode — it is generating episode seven and having it remember what happened in episodes one through six. StoryLab holds a series bible as durable state, runs arc review across the whole season, and gates each stage on a reviewer agent before anything reaches listeners. The series in this screenshot, Dungeon Hearts, is a LitRPG: HUDs, health bars, level bosses, permadeath. Long-horizon generated content, shipped to a real audience on a schedule.
LONG-HORIZON CONTINUITYSERIES-BIBLE STATEREVIEWER GATESMULTIMODAL: SCRIPT · TTS · VIDEOHUMAN-IN-THE-LOOP CONTROLS
LONG-HORIZON CONTINUITY · REVIEWER GATES · MULTIMODAL PRODUCTION
SHOT BOARD + THE CUT — 23 SHOTS, MULTIPLE TAKES PER BEAT, DRAG TO RE-ORDER, SUBMIT RE-CUT
Character identity has to survive 23 shots, and the timeline is the state — every re-cut re-derives what the agents must keep consistent. Takes are generated in parallel and scored before they enter the cut, so the human is making creative choices, not fixing broken frames.
This is the director's-chair model any team needs when generated content has to be authored, reviewed and shipped on a schedule: the agents do the volume, the human keeps the judgment, and the tool makes both legible.
SHOT-LEVEL AGENTSCHARACTER CONSISTENCYMULTI-TAKE GENERATIONHUMAN-STEERED CUTWEEKLY RELEASE CADENCE
SHOT-LEVEL AGENTS · CHARACTER CONSISTENCY · HUMAN-STEERED CUT
MAIN HALL — ROOM NAVIGATION, IN-CHARACTER SUSPECT DIALOGUE, LIVE EVIDENCE DISCOVERY
Nobody asked me to build this.
Every suspect is an agent with private knowledge, a motive to lie, and a story that must survive free-text interrogation — while the mystery stays solvable no matter what the player types.
World, agent architecture, React front end, production ops: all mine. This is me unsupervised, for users not obligated to be kind.
HIDDEN-KNOWLEDGE AGENTSDECEPTION MODELINGSOLVABILITY GUARANTEESFULL STACK, ONE PERSONSHIPPED & LIVE
▶ PLAY IT YOURSELF — AISHEWROTE.COMHIDDEN-KNOWLEDGE AGENTS · DECEPTION MODELING · SOLVABILITY GUARANTEES
MIXED-MEDIA FEED — INTEREST TAGS, PER-POST REACTION SIGNALS, CREATOR AND AGENT CONTENT SIDE BY SIDE
The feed is the reason the eval work matters. Every reaction here is a live quality signal, and those signals are exactly what Judge Dredd correlates against offline eval scores to find which dimensions actually predict engagement. It is also the hardest surface to get right: mixed media, wildly different formats per post, generated and human content in the same grid, and moderation that has to hold at 20M+ MAU without flattening what makes the content worth posting.
MULTIMODAL FEEDENGAGEMENT AS EVAL SIGNALMODERATION AT SCALE20M+ MAU
MULTIMODAL FEED · ENGAGEMENT AS EVAL SIGNAL · MODERATION AT SCALE
02 // THE SCORECARD
REQUIREMENT MATCH ANALYSIS
// TARGET · PRINCIPAL ENGINEER — APPLIED AI PRODUCTS, ENTERPRISE WORKFLOWS
// A REAL PUBLISHED REQUIREMENT LIST FOR THIS ROLE, SCORED AGAINST MY RECORD — INCLUDING THE ONES I MISS.
8DIRECT4ADJACENT
STRONGEST FOR
Agent architecture, evaluation infrastructure, and taking probabilistic systems from an ambiguous brief to production at scale.
WHAT I WOULD BE LEARNING
An enterprise permission model, front-end depth, and an enterprise customer in the room. Each of those rows says so in its own words.
Distinguished Engineer, AI Entertainment at Character.ai: I own technical and product strategy for agentic systems at 20M+ MAU, and I am still the author of the systems that run there — AgentX, the SDK, the eval wall.
Twenty years underneath it: Coinbase, Netflix, Google Cloud, VMware. Product judgment was learned by shipping and reading telemetry, not by writing PRDs.
AI She Wrote is the whole stack by one person — world, agent architecture, web front end, production ops. c.ai/studio is a production-floor UI over agent backends. Underneath: Go services, Spanner and ClickHouse data layers, systems at millions of QPS.
Enterprise integrations are the older chapter — at VMware I tech-led datastores, stream processing, identity and air-gapped enterprise readiness from zero to GA in six months.
Backend is where the depth is: Go for every agent runtime, Python for post-training and evals. I ship front ends in React and TypeScript — this site, AI She Wrote, the studio tooling — and I am fluent, not the most senior front-end engineer in the room.
If the role's hardest problems are in the browser, that is my ramp. If they are in the system behind it, that is home.
At Coinbase I set the company-wide Datastores strategy that teams I did not run had to live inside. At Character.ai the org-wide AI tooling standards are mine, which is mostly a persuasion problem rather than a technical one. Both are influence across an organisation, at company scale.
The customer half is the honest gap: my recent customers are players and internal engineering teams, not enterprise buyers. The last time I sat across from an enterprise design partner was VMware. I know how to do it; I have not done it lately.
This is the requirement I would point at first. JudgeJudy is the offline harness; Judge Dredd pulls Amplitude and Statsig telemetry over MCP, correlates eval scores against what users actually do, and only the dimensions that predict engagement become release gates.
Every AgentX release clears that wall before a user sees it. Uncertainty becomes a measured gate, not a meeting.
I have owned systems where a mistake is a headline: wallet infrastructure processing billions a day at Coinbase, entertainment infrastructure at 200M+ members, and org-wide AI safety guidelines and sandbox isolation enforced at deployment. At VMware I built identity and air-gapped readiness for enterprise customers.
What I have not done recently is design the permission model for an enterprise AI product — connectors, tenancy, data boundaries. I know what it costs to get it wrong; I would be learning the specific shape.
Each began as a prototype nobody had scheduled and is now load-bearing. I scope it, build it, ship it, and own the pager afterwards.
The decisions behind them — durable world state over a single long-context pass, a trained judge over human review, pairwise preference over absolute scores — are in the evidence above, with what each one cost.
AgentX needed a runtime. Instead of a one-off I built claude-agent-sdk-go, and five production systems now run on it — including Jeeves, an onboarding agent with a different goal and a different tool registry. At Coinbase, ClickHouse and CockroachDB went from one team's need to company-wide paved paths that outlived my tenure.
The call is the same each time: build the specific thing fast, and generalise only once a second workload proves the shape.
Google Cloud: a Spanner optimisation cutting latency up to 80%, quota management rebuilt at millions of QPS. Netflix: the self-service datastores platform at 200M+ members. Coinbase: seven datastore engines, an observability org built from scratch and a vendor replaced in-house at 90% savings, wallet infrastructure processing billions daily.
Character.ai: the WarpStream migration behind 5 GB/sec of multimodal ingest, on the record in the Confluent interview below.
A planner decomposes the request, executors generate image, video and audio, reviewer agents gate every stage on trained eval scores. I post-train open models with SFT and RL for the agentic use cases where a frontier model is too slow or too expensive per call, and the judge itself is a preference-trained 8B VLM calibrated against human annotation.
Reliability of probabilistic systems is not a bullet here. It is the job.
c.ai/studio is a workflow product: a shot board and cut timeline where a production team steers agents on a weekly release schedule. Judge Dredd is an insight product: which quality dimensions actually predict engagement. At Coinbase I built the observability org whose product was making complex systems legible to engineers.
All of that was for internal users or players. I have not shipped an analytics product to an enterprise customer, and I would rather say so.
I grew Datastores and Observability at Coinbase to 20+ engineers and set the company-wide technical strategy they executed. At Character.ai I set the org-wide standards for how agentic systems are built, evaluated and shipped, and the SDK is the pattern other engineers build on.
Every decision above is tied to a number — latency, cost, engagement, release regressions — because that is the only way I know to argue for one.
Every line above is a published qualification from a real posting for this role, answered with what actually exists rather than a list I picked to flatter myself. 8 are a direct hit. 4 I reach by a different route, and those rows say so in their own words rather than hiding behind a checkmark. If a principal who tells you what he cannot do is useful to you, the direct line is at the bottom of this page.
03 // THE TOOLING
OPEN SOURCE // AGENT PRIMITIVES
65 STARS · 3 REPOS · MIT
// THE INFRASTRUCTURE LAYER UNDER THE AGENTS, BUILT IN THE OPEN SO OTHER TEAMS DO NOT REBUILD IT.
04 // THE CAREER
THE RECORD // MEMORY SEQUENCES
0/6 OPEN
// TWENTY YEARS IN SIX SEQUENCES. 65 GITHUB STARS, TWO ORGS BUILT, AND A PAGER I STILL CARRY.
30+ agentic turns, 6 specialized tools, no human in the loop after the prompt. Production-grade, observable, reliable at 20M+ MAU. Jeeves, an adaptive onboarding agent, runs on the same SDK.
MULTI-AGENT · GO · TRAINED EVALS▸ OPEN THE SEQUENCE
Judge Dredd correlates eval scores against production telemetry over MCP, finds which quality dimensions actually predict engagement, and feeds validated gates back into the runtime automatically. The judge underneath is a fine-tuned Qwen3-VL trained on pairwise preference with Bradley-Terry loss — dataset curation, training, calibration against human annotations, and the retraining loop, in Python.
PREFERENCE-TRAINED VLM · PYTHON · MCP▸ OPEN THE SEQUENCE
Streaming, tool registries, hooks, MCP support, bounded goroutine pools. Gremlord keeps Claude Code exactly as shipped and swaps what sits behind it — any provider, every token metered, hard-stop budgets — in a static Go binary with no daemon.
GO · MCP · OPEN SOURCE (MIT)▸ OPEN THE SEQUENCE
Seven engines in production. Replaced an external observability vendor in-house at 90% cost savings. Architected Wallet-as-a-Service infrastructure processing billions in daily trading volume.
CLICKHOUSE · COCKROACHDB · TEAM OF 20+▸ OPEN THE SEQUENCE
Saved engineering 10–20 dev hours per week across the org. Contributed to Netflix OSS (Hollow — delta-based in-memory sync). Entertainment infrastructure at a scale where a bad deploy is a headline.
OSS CONTRIBUTOR · 200M+ MEMBERS▸ OPEN THE SEQUENCE
Rebuilt GCP Networking Quota Management at millions of QPS. At VMware: datastores, stream processing, identity, and air-gapped enterprise readiness. Foundation laid as an officer in the IDF Intelligence Corps.
SPANNER · MILLIONS OF QPS▸ OPEN THE SEQUENCE
05 // ON THE RECORD
TRANSMISSION // ON THE RECORD
● 2 RECORDINGS RECOVERED
Why Your Video Judge Rewards Slop — and How to Fix It
A generated clip where the character freezes for four seconds can still score well — the judge rewarded gloss instead of what actually happened. Video breaks the text eval playbook: temporal consistency, shot continuity, and story coherence don't show up in a CLIP score, and a team watching clips on Friday doesn't scale.
The fix that stuck: stop scoring, start comparing. Pairwise preference — is B a better story than A? — trained into a Qwen3-VL judge with Bradley-Terry loss on real and deliberately broken footage, calibrated against human scores, running as a CI regression gate before any user sees a release.
PAIRWISE PREFERENCEQWEN3-VL JUDGEBRADLEY-TERRY LOSSCI REGRESSION GATE
// EVERY TEAM SHIPPING GENERATED CONTENT TO USERS HITS THIS THE WEEK THEY GO LIVE.
PODCAST · INTERVIEWED BY CONFLUENT
Harnessing Data at Scale — Character.ai's Transition to WarpStream
The streaming backbone under everything above: 5 GB/sec of multimodal event data feeding the agents, the evals, and the telemetry loops. Confluent invited me on to walk through why we moved and what it cost us not to.
WARPSTREAM5 GB/SEC INGESTCOST-EFFICIENT STREAMING
▸ READ THE ENGINEERING BLOG POST06 // THE FILE
DATABASE // CV
OPERATIVE
BRIL, M.
Distinguished Engineer, AI Entertainment
SAN JOSE, CA
STATUS · ACTIVE
CLEARANCE · 20 YRS IN PRODUCTION
STATUS · ACTIVE
CLEARANCE · 20 YRS IN PRODUCTION
2024 — PRESENT
DISTINGUISHED ENGINEER, AI ENTERTAINMENT
Character.ai
Technical and product strategy for agentic AI across Character.ai's entertainment products at 20M+ MAU. AgentX, the agent SDK, post-trained models, and the eval wall gating every release.
2022 — 2024
SENIOR STAFF ENGINEER
Coinbase
Company-wide Datastores strategy; two orgs grown to 20+ engineers; 90% observability cost savings; billions in daily trading volume.
2021 — 2022
SENIOR SOFTWARE ENGINEER
Netflix
Self-Service Datastores Platform and the Data Abstractions team, 200M+ members. Netflix OSS contributor (Hollow).
2019 — 2021
SENIOR SOFTWARE ENGINEER
Google Cloud
Cloud Spanner optimization cutting latency up to 80%; GCP Networking Quota Management rebuilt at millions of QPS.
2015 — 2019
STAFF / TECH LEAD
VMware (acq. Arkin)
Tech-led a 6-person team from zero to GA in 6 months: datastores, stream processing, identity, air-gapped enterprise readiness.
2008 — 2012
BSc, EE & COMMUNICATIONS
HIT, Israel
Graduated while serving as an officer in the IDF Intelligence Corps.
// DIRECT LINE
INITIATE CONTACT
Everything in this archive is already running in production. The only open question is what I build next, and who I build it with.
Email me with the subject line FIRST 90 DAYS and I'll reply within 24 hours with a one-page plan for your first 90 days of agentic systems.
© 2026 MAOR BRIL · SAN JOSE, CA · MEMORY ARCHIVE v20.26
ALL SEQUENCES BATTLE-TESTED IN PRODUCTION
ALL SEQUENCES BATTLE-TESTED IN PRODUCTION