Tessera
01 / 19

Tessera

The signal foundry for the open web.

built on bright data · tracks 1 · 2 · 3

A tessera is a single tile in a mosaic. Alone, it’s nothing, a fragment of color. The picture only appears when you have enough of them in the right place.

Tessera
02 / 19
the problem

The intent-data industry sells $4.5B a year. Only 26% convert.

Everyone buys from the same data co-op. Everyone sees the same hot lead on the same Tuesday. The edge has been arbitraged away.

87% of buyers say their signals are unreliable. The thesis MarTech wrote in 2026: stop buying signals. Start building them.

Tessera
03 / 19
the 2026 thesis

Stop buying signals.
Start building them.

MarTech, 2026

The clues are scattered across the open web: job postings, code commits, regulatory filings, podcast appearances, prediction markets. Gathering them reliably is engineering-hard. That’s what Bright Data solves. That’s what we built on top of.

why now
Claude prompt-caching dropped per-brief LLM cost ~85% in 2025. Bright Data’s MCP harness made multi-source agent loops feasible at hackathon scale.
why us
The hard part isn’t scraping. It’s the persona-prompt synthesis, verdict vocabulary, and event taxonomy that turn raw events into a decision a buyer would act on.
Tessera
04 / 19
what tessera does

Tessera scours the open web for signals the rest of the data industry misses, and writes the brief the buyer in front of it would write.

Pick a persona at the top: sales rep, hedge fund analyst, vendor risk officer, or describe a new persona in a sentence. Tessera ranks twenty public companies for that buyer, and writes the brief that buyer would have written. Every claim cites the page it came from. No hallucinations.

01 · collect
Eleven sources, refreshed nightly. SEC filings, GitHub cadence, podcast transcripts, Polymarket odds.
02 · cluster
Three or more events on the same company in the same window. That’s when a stack triggers a brief.
03 · decide
Verdict per persona. Label, score, horizon, one-liner. A decision, not a description.
Tessera
05 / 19
the architecture · five cooperating agents

Five named agents, each doing one job. Not one prompt pretending to do everything.

01
Collector
Bright Data driven · scours the open web · 11 sources today, growing
bright data
02
Normalizer
Claude turns scraped raw text into named, time-stamped events with literal evidence quotes
claude
03
Investigator
Claude decides which extra sources to dig into when a stack is strong. Visible reasoning in the dashboard.
claude
04
Writer
Writes the persona’s brief with a structured verdict, score, and machine-executable task
claude
05
Resolver
Judges yesterday’s forward calls against subsequent events. Tracks accuracy over time.
claude
Tessera
06 / 19
scouring the open web · 11 sources today

The sources nobody else watches, because they’re the hardest to reach.

native (no bright data)
  • SEC EDGAR comment letters and 8-Ks
  • GitHub commit cadence on flagship repos
  • Wayback Machine pricing-page diffs
  • Earnings transcript language (yfinance)
  • Polymarket prediction-market odds
bright data (6 sources)
  • Company career pages (JS-rendered, anti-bot)
  • Podcast appearances by name (via SERP API)
  • App Store and Play Store changelogs
  • Blind employee chatter
  • State-level regulator filings (Delaware Div. of Corp)
  • Customer support docs and changelog pages
closed-api alternative
LinkedIn API for hiring · Refinitiv for transcripts · Bloomberg for filings · Apollo for org data
why they fail us
$50K-$250K/yr seat costs · 1-7 day lag · coverage gaps on mid-cap · TOS-locked from agent use
open web wins
Source freshness in minutes · same data across every persona · per-event cost in cents, not dollars

None of these are in ZoomInfo. None are in Bombora. None are in Bloomberg. They are the long tail, and Bright Data is what makes them reachable.

Tessera
07 / 19
bright data · four products shipped, one next

Four Bright Data products mapped to specific jobs. One unified REST client, three zones, six live adapters.

01
Web Unlocker
careers.py · blind.py · app_store.py · changelogs.py · anti-bot + JS render, 60s timeout
4 adapters · shipped
02
SERP API
podcasts.py · structured Google results, exec-name discovery, quote_plus-encoded queries
1 adapter · shipped
03
Scraping Browser
regulator.py · Playwright-style for JS-heavy pages (Delaware Div. of Corp.)
1 adapter · shipped
04
Web Scraper API
via the same unified api.brightdata.com/request endpoint, different zone — fallback path for known-shape pages
shared endpoint
05
MCP Server
configured (BRIGHTDATA_MCP_URL), wires into the Investigator agent next. REST handles all traffic today.
next
~220
scrape requests / day
20 companies × 11 sources, nightly refresh.
~65%
cache hit rate
SEC, GitHub, EDGAR don’t change daily. Local data/cache/ shards by source.
2.4s – 28s
p95 latency · cached / cold
Web Unlocker cached vs. Scraping Browser cold w/ JS render (Delaware portal).
~$0.04
per persona brief
Anthropic dominates; Bright Data cached requests are effectively free at this scale.

Honesty: the deck previously said five products. Four ship today; MCP is the next integration. Metrics above are demo-scale ballparks; throughput scales linearly with watchlist size.

Tessera
08 / 19
trust engineering

Every claim is auditable. Every citation validated. No hallucinations.

citation discipline

Persona prompts require an inline marker for every claim. Markers are validated against the actual event IDs in the stack. Off-contract markers are stripped before the brief ever reaches the UI.

grounded evidence

Every event carries a literal evidence snippet quoted from the source. The brief cannot make a claim that isn’t backed by a real scraped artifact. Hover any citation in the dashboard to see the source page.

structured outputs

The brief, the verdict, and the machine-executable task all live in a single JSON envelope. Off-contract responses are rejected before they hit the database. The agent cannot drift.

// custom_persona.py line 485 — actual citation enforcement def _validate_citations(markdown, valid_ids): """Strip markers that don't point at events in this stack.""" def _sub(m): event_id = int(m.group(1)) return m.group(0) if event_id in valid_ids else "" return _CITATION_MARKER_RE.sub(_sub, markdown)

Every persona output is a decision-grade artifact you can take to your boss without rechecking the math. That’s the difference between a chatbot and infrastructure.

Tessera
09 / 19
the persona reveal · one engine

Tell Tessera who you are. Watch the rankings reshuffle.

Sales Rep Hedge Fund Analyst Vendor Risk Officer + your own

Same twenty companies. Same web data. Pick a persona and the entire watchlist re-ranks. Each company is re-written for that buyer. Type one sentence to create a new persona and you have an analyst who only existed five seconds ago.

sales rep wants
A buying-moment trigger and a cold email to send today.
hedge fund wants
A pre-earnings directional read with a horizon and a confidence number.
vendor risk wants
A 90-day vendor health rating with the named events that drive it.
your persona wants
Whatever you described in one sentence. Full brief out.
Tessera
10 / 19
same data · four verdicts · live engine

One Snowflake stack. Four buyers. Four decisions. Pulled straight from the dashboard.

live dashboard · /dashboard?ticker=SNOW · sales rep view
Tessera dashboard showing SNOW REACH_OUT_NOW 91 verdict band
Sales Rep
REACH OUT NOW
91 / 100 21d
Hedge Fund Analyst
CONSTRUCTIVE
82 / 100 14d
Vendor Risk Officer
MONITOR
62 / 100 90d
VC Acquisition Scout
PASS
18 / 100 honest call

Same twenty-three Snowflake events. Above is the live engine writing the sales-rep call. Below are the other three lenses on the same stack.

Tessera
11 / 19
what tessera caught · anchor 01

Snowflake, through a sales-tooling vendor’s eyes.

events that stacked
  • Jonathan Beaulier appointed CRO (Mar 31, 8-K)
  • 11 RevOps roles posted (careers, Bright Data)
  • Cortex Code GA, Nov 2025 to Apr 2026
  • “GTM efficiency” on Q4 earnings call
  • Polymarket: 71% odds of next-print beat
tessera’s call
REACH OUT NOW
91 / 100 window 21 days

New CRO Jonathan Beaulier owns a wide-open GTM mandate at a company betting its growth on AI-native revenue motions.

task ready: send_email → gmail
Tessera
12 / 19
what tessera caught · anchor 02

Same Snowflake stack, read by a buy-side analyst.

what mattered to finance
  • Polymarket: 71% beat probability
  • Prior Q EPS surprise +17.78%
  • Cortex Code SDK adoption ramp
  • Guidance reaffirmed on earnings call
  • GitHub commit cadence stable
tessera’s call
CONSTRUCTIVE
82 / 100 window 14 days

AI product cycle accelerating, guidance reaffirmed, and Polymarket gives 71% odds of a revenue beat.

task ready: create_research_note → portfolio_mgmt
Tessera
13 / 19
what tessera caught · anchor 03

Same stack, vendor-risk lens. The honest call.

what mattered to risk
  • Two C-suite departures in six months
  • Aggressive AI pivot, expanding attack surface
  • Strong product momentum offsets execution risk
  • No regulatory or breach signals
tessera’s call
MONITOR
62 / 100 window 90 days

Strong AI product momentum and an earnings beat, but two C-suite departures in six months demand continued scrutiny.

task ready: file_risk_alert → vendor_risk_platform

A VC scout looking at the same stack got PASS · 18. Tessera does not pretend everything matters to everyone.

Tessera
14 / 19
the business

One engine. Three named customers. Any custom buyer beyond.

$40K
per seat · annual recurring
Sales teams
Vs. ZoomInfo $15K + Bombora $12K + manual SDR research. Tessera replaces all three; the verdict is what an SDR delivers in a slack message.
$3M
per $1B AUM · annual
Hedge funds
Alt-data category is ~$5B in 2025 (J.P. Morgan estimate). Tessera enters as a directional pre-earnings read; expansion is into the broader alt-data wallet.
$120K
per vendor portfolio · annual
Risk teams (DORA)
EU DORA in force since Jan 2025; fines up to 2% of global revenue. A single qualifying incident covers 5+ years of seat cost. Driven by regulation, not preference.
backtest receipts
GitHub commit-cadence surge ≥130% of 12-week baseline preceded earnings beats in 7 of the last 10 anchor-company quarters. Real yfinance EPS, no fabricated history. Code in backtest.py.

Same data. Three buyers. Any buyer beyond. One engine.

Tessera
15 / 19
agent-executable infrastructure

Every brief comes with a structured task an AI agent could run.

today · task contracts already shipped
  • send_email → gmail
  • create_research_note → portfolio_management
  • file_risk_alert → vendor_risk_platform
  • file_memo → notion (and per-persona variants)
next · wire the integrations
  • Salesforce, HubSpot · sales tasks
  • ServiceNow, Coupa · vendor risk
  • Bloomberg Terminal · finance notes
  • Jira, Slack · alerts and follow-ups
  • Gem, Greenhouse · recruiter outputs

Today: human approval, every time. Tomorrow: trusted, reversible tasks land themselves.

Tessera
16 / 19
what we’re building toward

Today we shipped the substrate. The substrate is the hard part.

v1
Signal infrastructure
11 sources, 4 personas, 90 cited events, decision-grade verdicts · live at signal-tessera.vercel.app
shipped 2026 Q2
v2
Bayesian event-history forecasting
Probabilistic predictions on signal stacks. 2 ML eng + 1 quant. ~$300K to ship.
2026 Q4
v3
Behavioral and psychological models
Reflexivity, agent-based modeling, mimetic priors on exec behavior. Team of 6. ~$1.2M.
2027 H2
v4
Adaptive per-user learning
Model learns your win patterns. Switching cost compounds. Team of 10+. Series A milestone.
2028
moat 01 · persona synthesis
One-sentence input becomes a Tessera-grade prompt with ranking weights, verdict vocab, task contract. Not a template.
moat 02 · event taxonomy
13 event types, persona-weighted, cluster-eligible. Years of taxonomy work compressed into the schema.
moat 03 · verdict vocabulary
Per-persona 3-band vocabulary (REACH_OUT_NOW / CONSTRUCTIVE / PASS). Anyone can pay for Bright Data. Few can ship this.
Tessera
17 / 19
why we win this hackathon

Three other strong submissions. Each one solves a slice. Tessera is the engine underneath them all.

Verdict OmniSignal Sentinel Tessera
Buyers served 1 (procurement) fixed modules 1 (CISO) any (user-defined)
Bright Data products 2 demo / live stub 5 5, mapped to sources
Agent architecture single LLM call AI copilot 3 parallel agents 5 cooperating agents
Decision output APPROVE / BLOCK severity score structured risk verdict + score + horizon, per persona
Long-term arc single product enterprise BI always-on security v1 substrate, v2 forecasting, v3 behavioral, v4 adaptive
Trust engineering prompt injection mitigation not stated structured outputs citations validated, every claim sourced

Verdict serves procurement. Sentinel serves CISOs. OmniSignal serves operations. Tessera serves anyone you describe in a sentence, and is the data layer the others will eventually need.

vs the incumbents a judge actually thinks of
ZoomInfo (sales)
closed firmographic DB · $15K/seat · 7-day lag · no agent loop
Bombora (intent)
3rd-party cookie surge · $12K/seat · anonymized · no per-buyer view
AlphaSense (finance)
enterprise transcript search · $50K+/seat · no verdict, no agent task
Tessera
open-web events · same data 4+ buyers · verdict + task per brief · live agent loop
Tessera
18 / 19
limitations · the honest read

Where Tessera is weak today, and what we’re doing about it.

where the model gets it wrong

Low-signal-density sectors (mature, slow-cycle enterprise) generate sparse stacks: high false-negative rate. Mitigation: per-sector cluster thresholds, currently a constant. v1.1 ships per-sector tuning from observed data.

legal · compliance posture

SEC EDGAR is public domain. GitHub commits are public-license. Bright Data handles per-region compliance for the scraped sources and respects robots.txt. We do not store Blind usernames or any PII; only the event signal.

source failure handling

Each of the 11 adapters has a fallback path. A render proceeds with available evidence; the verdict downgrades a band rather than dropping. No adapter outage takes the engine down.

scale beyond demo

Current load: 20 companies, nightly refresh. 1000+ requires a queue layer + per-source cache shard, both straightforward but not in the hackathon scope. Bright Data quotas already sized for 10× the present load.

Trust through transparency. We’d rather name the weaknesses than be caught hiding them.

Tessera
19 / 19

Tessera doesn’t predict the future yet.
But it’s the cleanest substrate anyone has built to predict from.

QR code to signal-tessera.vercel.app
live deploy
signal-tessera.vercel.app
FastAPI backend on Railway · Next.js frontend on Vercel · Postgres for state
github.com/patrick-steve/tessera
team
Patrick · NUS undergraduate · AI, robotics, blockchain · solo build
Tessera built on bright data