01
A method for small-sample research
Ten real voices,
a hundred answers.
Turning a hard-to-recruit sample into a validated respondent panel — without pretending the synthetic ones are real.
Prepared for Pramana Yoga
Politeknik Negeri Malang
Elux Space · 2026
02
The problem
Good research needs the right respondents.
The right respondents are hard to find.
The need
Validate with a real sample
A study is only credible if it's tested on people who genuinely match the target profile — the exact segment your research is about.
The reality
Narrow criteria don't scale
The tighter the profile — location, age, background, skill — the longer recruiting takes. Weeks of effort, and still only a handful reply.
The consequence
The study stalls
You either wait indefinitely, or ship on a sample too small to defend. Neither is a good answer for real research.
Synthetic Respondent PlatformThe problem
03
A concrete example
You needed 100.
You found 10.
The target: 100 respondents in Malang — vocational high school, second year, one year of C++ programming.
📍 Malang, East Java
🎓 SMK · 2nd year
⌨️ 1 year of C++
After weeks of recruiting, only 10 people matched every criterion. That's the wall this platform is built to get over.
● 10 real, found
○ 90 missing
Synthetic Respondent PlatformThe gap
04
The idea
Don't fake the data.
Amplify it.
Those 10 people aren't just 10 data points — together they carry a pattern: how this kind of person thinks, decides, and answers.
Interview them deeply enough to capture that pattern, and you can extend it into a larger panel — where every synthetic respondent is an echo of a real one, never an invention.
The 10 real answers stay real and labeled. The 90 synthetic ones are clearly marked as synthetic — and measured against reality before you trust them.
Synthetic Respondent PlatformThe idea
05
How it works
Four steps, from a few to many.
1
Listen
A deep calibration questionnaire goes to the 10 real people — capturing how they think and behave, not just their demographics.
2
Find the pattern
A capable AI reads all 10 and extracts a behavioral model: traits, tendencies, and the axes along which people vary. You review and approve it.
3
Generate the many
90 synthetic respondents are created — each a coherent persona sampled from the pattern, answering your actual survey in Bahasa Indonesia.
4
Prove it
Two real people are held back and used to score the synthetic answers against reality — giving every study an honest accuracy grade.
The researcher stays in control. Nothing is generated until you've seen and approved the pattern the AI extracted from your real respondents.
Synthetic Respondent PlatformThe pipeline
06
What goes in, what comes out
You bring the truth.
The platform brings the scale.
You provide — the input
- Target criteria — the exact profile you're studying
- Calibration questionnaire — deep questions on behaviour & reasoning
- Target survey — the questionnaire you actually want answered
- 10 real responses — collected via a shareable form link
The platform returns — the output
- 90 synthetic respondents — coherent personas with full answers
- One combined dataset — every row labelled real or synthetic
- Aggregate view — distributions, real-vs-synthetic overlay
- Accuracy grade — how well the synthetic panel tracks reality
Real and synthetic data are never silently mixed. The label travels with every export — CSV, JSON, or dashboard — so any downstream analysis stays honest.
Synthetic Respondent PlatformInput & output
07
Technical architecture
One small system on Cloudflare.
Synthetic Respondent PlatformArchitecture
08
The engine
Two models, two jobs.
A smart model for the rare, hard job. A cheap model for the bulk work. That split is what keeps quality high and cost low.
Extraction · the analyst
GPT-5.5
Reads deeply, thinks carefully.
Runs: once per study
Job: turn 10 rich interviews into one behavioural pattern
Why premium: the pattern decides everything downstream — quality here is worth the cents
1×
Generation · the workforce
DeepSeek V3
Fast, cheap, coherent at volume.
Runs: 90 times per study
Job: each call = one full persona answering the whole survey
Why cheap works: the hard thinking is already done — this is disciplined role-play at scale
90×
Synthetic Respondent PlatformThe engine
09
What it costs
Priced per token.
Measured per study.
| Stage | Model | Input tok | Output tok | Est. cost |
|---|---|---|---|---|
| Extraction (1×) | GPT-5.5 | ~30 K | ~15 K | ~$0.30 |
| Generation (90×) | DeepSeek V3 | ~450 K | ~135 K | ~$0.27 |
| Validation (holdout) | GPT-5.5 + V3 | ~25 K | ~12 K | ~$0.13 |
| Full study — 100 respondents | ~505 K | ~162 K | ≈ $0.70 |
Token counts are planning estimates for a ~40-question study. DeepSeek V3 ≈ $0.27 / M input, $1.10 / M output; GPT-5.5 pricing to confirm at build. Prompt caching on the shared pattern lowers generation cost further.
Headline
≈$0.70
to turn 10 real respondents into a labelled panel of 100.
At 20 studies / month, that's roughly $14 / month in AI spend — the recruiting cost of a single respondent.
Synthetic Respondent PlatformCost
10
Quality of the output
Every study
grades itself.
Before generating, the platform quietly sets aside 2 of the 10 real respondents. It then predicts their survey answers from their profile — and compares against what they actually said.
Closed questions are scored by match rate; open answers by an independent AI judge. The result is a single, honest grade attached to every export.
The study grade
A
Tracks reality well — safe for directional decisions.
≥ 0.80
B
Usable with care — review the low-confidence questions.
0.65–0.79
C
Weak — recruit more real people or fix the questionnaire.
0.50–0.64
D
Do not use the synthetic data from this study.
< 0.50
A 2-person holdout is a coarse check — it reliably catches broken studies, but does not certify a perfect one.
Synthetic Respondent PlatformQuality
11
Constraints & honest limits
What this is — and isn't.
- Directional, not a replacement. Great for pretesting surveys, sizing sentiment, and early hypotheses — not a substitute for full fieldwork.
- The real sample sets the ceiling. Eight to ten people can only carry so much signal. A minimum real sample is enforced before generation unlocks.
- It inherits model bias. An AI's idea of "a student" can drift from a real one — countered by evidence-tracing and the holdout grade.
- Synthetic never hides. No export, chart, or screenshot ever presents synthetic data without its label. This is structural, not a setting.
- Quality depends on the questions. A shallow calibration questionnaire yields a shallow pattern. The depth of the interview is the real lever.
- No pure fiction. The platform will not generate respondents without a real sample to ground them. There is no "invent from scratch" mode.
Sold honestly, this is a tool that amplifies a small real sample — and says so, on every deliverable, alongside a grade for how far the amplification can be trusted.
Synthetic Respondent PlatformConstraints
12
Infrastructure & its limits
Built to be small, cheap, and durable.
Workers · compute + UI
D1 · database
Workflows · long jobs
Access · team auth
Turnstile · bot defense
| Limitation | Why it exists | How it's handled |
|---|---|---|
| Worker request time | A minutes-long 90-call batch can't run in one request | Workflows own all AI batch work; the UI polls job status |
| LLM rate limits | Providers cap requests per minute | Concurrency capped at 5, with automatic back-off & retries |
| Malformed AI output | Models occasionally break the JSON schema | Validate → retry up to 3× → flag; the job never hangs |
| Database size | D1 has a per-database ceiling | Non-issue — a full study is only a few megabytes |
The whole system fits inside Cloudflare's free-to-modest tiers. No servers to run, nothing to keep warm — it costs almost nothing when idle, and scales by study, not by month.
Synthetic Respondent PlatformInfrastructure
13
Where this goes next
From 10 respondents
to a study you can defend.
The natural first pilot is your own Malang C++ study — ten real respondents, expanded, validated, and graded end to end.
Let's talk about the pilot, Pak Pramana.
Elux Space · 2026